1008-00005 · Lumos Analytics Inc · Pleadings

declaration-fong.docx DOCUMENT Litigation

20 / 35 ← → ✕

SUPERIOR COURT OF THE STATE OF CALIFORNIA

COUNTY OF SANTA CLARA

I, Rachel Fong, declare as follows:

Introduction and Personal Background

1. My name is Rachel Fong. I am over the age of eighteen years and am competent to testify to the facts set forth in this declaration. I have personal knowledge of each of the matters stated herein and, if called as a witness, could and would competently testify thereto.

2. I currently serve as Vice President of Engineering at Lumos Analytics Inc. ("Lumos" or "Lumos AI"), a Delaware-incorporated, California-headquartered software-as-a-service company specializing in artificial-intelligence-driven financial analytics for institutional asset managers. Lumos's principal place of business is located at 4700 Orchard Parkway, Suite 300, San Jose, California 95134. In my capacity as Vice President of Engineering, I am responsible for the strategic direction, development, and maintenance of all of Lumos's proprietary technology platforms, including the flagship LumosEdge product described in detail below.

3. I have over eighteen years of professional experience in software engineering, machine-learning systems, and data-infrastructure design. Prior to joining Lumos, I held senior engineering leadership positions at two prominent enterprise software companies, where I led teams responsible for building large-scale data-processing and predictive-analytics platforms. At Lumos, I oversee the engineering organization, which is the company's largest department. As of January 2023, Lumos employs approximately 340 people across all functions, and the engineering department comprises approximately 165 of those employees, including machine-learning engineers, data scientists, software engineers, data-pipeline engineers, infrastructure engineers, and quality-assurance engineers. I report directly to Lumos's Chief Executive Officer, Priya Chandrasekaran.

4. I joined Lumos in April 2018, approximately one year after the company's founding, and have served continuously in roles of increasing responsibility since that time. Over the course of nearly five years with the company, I have developed deep and comprehensive familiarity with every aspect of Lumos's proprietary technology, including the specific trade secrets and proprietary materials at issue in this matter. During my tenure, I have been directly involved in the hiring, management, and supervision of all engineering personnel, including Defendants Rajiv Anand, Serena Mallick, and Derek Yun. I had direct familiarity with the work performed by each of these individuals during their respective tenures at Lumos, including the specific proprietary systems, codebases, datasets, and confidential business information to which they had access.

5. I make this declaration for the purpose of providing the Court with detailed technical information regarding the proprietary technology and trade secrets at issue in this matter, and to describe the findings of the internal IT security investigation conducted by the Lumos information-security team under my oversight, which identified the specific data-exfiltration events carried out by the individual Defendants in connection with their departures from the company. The facts set forth herein are based on my personal knowledge, my review of Lumos's business records maintained in the ordinary course of business, and information provided to me by members of the Lumos IT security team who report to me.

Lumos's Business and Proprietary Technology

6. Lumos Analytics Inc. was founded in approximately January 2017 by Priya Chandrasekaran and two co-founders with the mission of applying advanced machine-learning techniques to financial data to generate actionable, predictive analytics for institutional asset managers. Since its founding, Lumos has built and continuously refined its flagship software-as-a-service platform, branded "LumosEdge." The LumosEdge platform uses proprietary ensemble machine-learning models to ingest and analyze vast quantities of traditional financial data and alternative data, and to generate predictive signals that institutional clients — including hedge funds, mutual fund complexes, pension funds, and sovereign wealth funds — use for equity and fixed-income portfolio optimization. As of the end of fiscal year 2022, Lumos achieved annual recurring revenue ("ARR") of approximately $67 million, reflecting strong and accelerating adoption of the LumosEdge platform across the institutional asset-management industry. To date, Lumos has raised approximately $185 million in total venture capital funding through its Series D round of financing. Key funding milestones include a $40 million Series B round led by Ridgeline Ventures, which closed in August 2019, and a $72 million Series D round led by Northstar Growth Equity, which closed in June 2022. These investments reflect the substantial market value placed on Lumos's proprietary technology and competitive position.

7. Lumos's competitive advantage in the marketplace rests fundamentally on its proprietary technology — specifically, its machine-learning model architectures, feature-engineering libraries, data-ingestion and transformation pipelines, curated training datasets, and confidential strategic product plans. This technology has been developed over the course of approximately six years at enormous expense. The cumulative investment in the technology at issue represents thousands upon thousands of engineering hours and tens of millions of dollars in direct research-and-development expenditures, not including the substantial indirect costs of recruiting and retaining the specialized engineering talent necessary to build and maintain these systems. The technology is not publicly available, has never been published or open-sourced, and is closely guarded as among the company's most valuable assets.

8. Lumos maintains rigorous and multi-layered security protocols specifically designed to protect its proprietary technology and trade secrets from unauthorized access, use, and disclosure. These protocols include, but are not limited to: (a) role-based access controls ("RBAC") that restrict access to specific repositories, systems, and datasets based on an employee's job function and need-to-know; (b) mandatory multi-factor authentication ("MFA") for all access to internal systems, including code repositories, data warehouses, and cloud infrastructure; (c) encryption of all proprietary data and source code both at rest and in transit, using industry-standard AES-256 encryption; (d) network segmentation that isolates sensitive development and data environments from general corporate networks; (e) comprehensive logging and monitoring of all data-access events, file transfers, email attachments, and cloud-upload events through integrated data-loss prevention ("DLP"), endpoint detection and response ("EDR"), and cloud access security broker ("CASB") systems; and (f) requiring all employees who are granted access to proprietary technology or confidential business information to execute a Proprietary Information and Inventions Assignment Agreement ("PIIAA") as a condition of their employment.

9. The PIIAA executed by each Lumos engineering employee contains several critical provisions designed to protect the company's intellectual property. Specifically, each PIIAA includes: (a) an assignment-of-inventions clause, under which the employee assigns to Lumos all right, title, and interest in any invention, improvement, development, discovery, or trade secret conceived or developed during the employee's tenure and within the scope of the employee's employment or using Lumos resources; (b) a confidentiality and non-disclosure provision that prohibits the employee from disclosing, publishing, or otherwise making available to any third party any Lumos confidential or proprietary information, which provision expressly survives termination of employment for a period of three years; (c) a non-solicitation-of-employees clause that prohibits the employee from directly or indirectly soliciting, recruiting, or inducing any Lumos employee to leave the company for a period of twelve months following the employee's separation; and (d) a covenant not to use or disclose Lumos trade secrets at any time, whether during or after employment. The PIIAA further requires the employee to return all company materials — including all documents, files, devices, code, data, and any copies thereof — upon termination of employment for any reason, and to certify in writing that all company materials have been returned and that no copies have been retained. These restrictive covenant provisions are critical to protecting Lumos's intellectual property and maintaining the confidentiality of its trade secrets.

10. In my capacity as Vice President of Engineering, I was personally and directly involved in overseeing the development and ongoing maintenance of each of the five categories of proprietary technology described in the following paragraphs. I am intimately familiar with the technical details, competitive significance, and protective measures associated with each category.

Category 1: LumosEdge Ensemble ML Model Architecture

11. The LumosEdge ensemble machine-learning model architecture is the core technological engine of the LumosEdge platform. This architecture consists of a proprietary neural-network topology that combines multiple distinct model types — including deep neural networks, gradient-boosted decision trees, and proprietary attention-based transformer variants — into an ensemble framework specifically designed and optimized for predictive signal generation in equity and fixed-income portfolio contexts. The architecture incorporates proprietary hyperparameter configurations that have been carefully tuned through extensive experimentation and validation. The specific design choices embodied in the ensemble architecture — including the number and types of constituent models, the topology of each model, the methods by which model outputs are combined and weighted, the proprietary loss functions employed, and the precise hyperparameter values — collectively represent Lumos's most valuable and closely guarded technical asset. The ensemble architecture is what enables LumosEdge to generate the predictive signals that differentiate Lumos's offering from those of its competitors.

12. The LumosEdge ensemble model architecture was developed over a multi-year period, beginning in approximately mid-2018 and continuing through the present, by a core team of senior machine-learning engineers working under my supervision. Defendant Rajiv Anand was a central member of this core team. Anand served as Principal Engineer — ML Infrastructure at Lumos from March 2019 through September 30, 2022, and was one of the key architects of the LumosEdge signal-generation engine. In that role, Anand had deep, hands-on access to the complete ensemble model architecture, including all neural-network topology definitions, hyperparameter configuration files, model-training pipelines, and associated technical documentation. His contributions to the architecture were significant, but they were made as a Lumos employee, using Lumos resources, within the scope of his employment, and subject to the PIIAA he executed upon hire.

13. The ensemble model architecture, including all topology definitions and hyperparameter configurations, is not disclosed publicly in any form. It is not described in any published papers, blog posts, patent applications, or marketing materials. The architecture is stored exclusively on access-restricted internal code repositories that require both role-based authorization and multi-factor authentication to access. Access to the specific repositories containing the model architecture specifications is limited to a small group of senior ML engineers and engineering leadership. The architecture represents significant competitive value — based on my experience and knowledge of the relevant field, I estimate that recreating equivalent models of comparable performance from scratch, without access to the proprietary architecture, would require a well-resourced team of experienced ML engineers a minimum of two to three years of intensive research and development, at a cost of many millions of dollars.

14. The loss of confidentiality in the LumosEdge ensemble model architecture would allow a competitor to replicate the core signal-generation capabilities of the LumosEdge platform without incurring the same development time and cost that Lumos invested. Such a competitor would be able to stand up a functionally equivalent predictive-analytics engine in a dramatically compressed timeframe, enabling it to enter Lumos's core market and compete directly for Lumos's institutional clients on the basis of technology that Lumos developed at great expense.

Category 2: Proprietary Feature-Engineering Libraries

15. Lumos has developed and maintains a proprietary set of feature-engineering libraries containing 2,847 individually engineered features specifically designed for financial signal generation. In machine learning, "features" are the specific, mathematically defined transformations applied to raw input data to create the structured inputs that are fed into predictive models. Lumos's 2,847 proprietary features transform raw financial data — including price, volume, order-book, fundamental, macroeconomic, and sentiment data — as well as alternative data from non-traditional sources, into precisely calibrated inputs for the LumosEdge ensemble models. Each feature embodies a specific hypothesis about informative patterns in financial data, and each has been individually researched, designed, implemented, tested, and validated by Lumos's data science and engineering teams.

16. The feature-engineering library is the product of years of extensive research, experimentation, and iterative testing. The process of developing effective features for financial signal generation is one of the most time-intensive and intellectually demanding aspects of building predictive ML models for financial applications. It requires deep domain expertise in both quantitative finance and machine learning, combined with access to high-quality data and substantial computational resources. The specific features in Lumos's library, their mathematical formulations, the methods by which they are combined and selected for use in different model configurations, and the proprietary insights they embody about the structure of financial data, are all proprietary and confidential. The feature library is maintained in access-restricted internal repositories and is not disclosed publicly in any form.

17. Defendant Derek Yun served as Machine Learning Engineer II at Lumos from January 2021 through October 21, 2022. In that role, Yun had direct access to the proprietary feature-engineering libraries as part of his daily work. Yun's responsibilities included implementing new features, testing feature performance, and integrating features into the model-training pipeline. Through this work, Yun gained intimate knowledge of the feature library's contents, including the mathematical formulations, code implementations, and performance characteristics of a significant portion of the 2,847 engineered features.

18. Access to the complete set of 2,847 proprietary features would provide a competitor with an enormous head start in developing a competing predictive-analytics platform. Rather than investing the years of research and experimentation necessary to identify and validate effective features, a competitor in possession of Lumos's feature library could immediately deploy those features in its own models, capturing the value of Lumos's investment without bearing the cost. The feature-engineering process is, in my professional judgment, one of the most valuable and difficult-to-replicate components of any financial machine-learning system, and Lumos's library represents an asset of extraordinary competitive significance.

Category 3: DataStream ETL Pipeline Codebase

19. Lumos has developed a proprietary data-ingestion and transformation system known internally as "DataStream." DataStream is an extract, transform, and load ("ETL") pipeline consisting of a proprietary codebase and accompanying data-schema definitions that manage the end-to-end process of ingesting, cleaning, normalizing, enriching, and delivering data from 43 distinct alternative-data vendors into the LumosEdge platform. Alternative data — including satellite imagery analytics, web-scraping feeds, credit-card transaction aggregates, geolocation data, social-media sentiment data, and other non-traditional data sources — is a critical input to the LumosEdge models, and the DataStream pipeline is the system that makes this data usable.

20. The DataStream pipeline was built and maintained by a specialized data-pipelines engineering team within the Lumos engineering organization. Defendant Serena Mallick served as Senior Software Engineer — Data Pipelines at Lumos from June 2020 through October 14, 2022. During her tenure, Mallick was the primary engineer responsible for building and maintaining the DataStream system. She designed and implemented many of the vendor-specific integration modules, wrote core components of the data-transformation logic, and maintained the data-schema definitions that govern how data flows through the pipeline. Mallick had complete and unrestricted access to the entire DataStream codebase, all associated schema definitions, configuration files, and technical documentation.

21. The DataStream codebase includes proprietary logic for parsing, deduplicating, reconciling, and enriching data from each of the 43 vendor integrations. Each vendor integration represents significant engineering effort, as each vendor delivers data in its own format, with its own quirks, quality issues, delivery schedules, and API specifications. The DataStream codebase embodies proprietary knowledge about vendor-specific data quality issues — including specific fields that are unreliable, known data gaps, latency characteristics, and edge cases in vendor data formats — as well as the optimal transformation methodologies that Lumos developed through extensive testing to address these issues. The integrations with these 43 vendors took years of cumulative engineering effort to develop, test, validate, and maintain, and the resulting codebase represents a formidable barrier to entry for any competitor seeking to build a comparable data-ingestion capability.

22. The complete DataStream repository, including all source code, configuration files, data-schema definitions, test suites, and internal documentation, represents approximately 780 megabytes of data. This codebase is stored in a private, access-restricted Git repository hosted on Lumos's internal Aldermere Software instance. Access to the DataStream repository is limited to members of the data-pipelines team and senior engineering leadership, and all access is logged and monitored.

Category 4: Curated and Labeled Training Dataset

23. Lumos maintains a proprietary training dataset of approximately 14.6 terabytes of curated and labeled historical financial and alternative data. This dataset was assembled over the entire course of Lumos's operations, beginning in 2017, and represents the cumulative output of years of data acquisition, curation, cleaning, labeling, and annotation by Lumos's data science team. The dataset includes historical market data, fundamental data, macroeconomic indicators, and alternative data from dozens of sources, all of which have been processed, aligned, and structured into a format optimized for training the LumosEdge ensemble models.

24. Critically, the training dataset includes not only raw data that has been licensed from third-party data providers under agreements that contain strict restrictions on redistribution and sublicensing, but also substantial quantities of proprietary labels, annotations, and derived data that were created by Lumos's data science team through extensive manual and semi-automated processes. These proprietary labels and annotations — which encode expert judgments about market regimes, event classifications, signal quality, and other domain-specific concepts — are among the most valuable components of the dataset and are entirely the product of Lumos's own intellectual effort. The training dataset is essential for training and validating the LumosEdge models, and its quality and comprehensiveness are a significant competitive differentiator. The investment in data acquisition alone — including licensing fees paid to third-party data vendors over the past six years — amounts to tens of millions of dollars, and the investment in curation and labeling represents thousands of hours of specialized data-science work.

25. Defendants Anand and Yun each had access to portions of this training dataset through their respective work on the ML infrastructure and model-training pipelines. Anand, as Principal Engineer — ML Infrastructure, had broad access to the training-data infrastructure, including the systems used to store, version, and serve training data to model-training jobs. Yun, as Machine Learning Engineer II, had access to specific subsets of the training data used in the model-training experiments he conducted as part of his daily work. Both individuals had sufficient access to obtain significant portions of the proprietary training dataset, including the proprietary labels and annotations.

Category 5: Confidential Product Roadmap — LumosEdge 3.0

26. In addition to its existing proprietary technology, Lumos maintains confidential product roadmap documents that detail the company's strategic product direction for the period 2023 through 2025. These documents include detailed specifications for the as-yet-unreleased "LumosEdge 3.0" product version, which represents the next major evolution of the LumosEdge platform. The roadmap document is a 37-page PDF that contains, among other things: planned new features and capabilities, detailed technical specifications for new model architectures and data-processing systems, competitive analysis of Lumos's market position relative to identified competitors, go-to-market timelines and launch strategies, target customer segments, and pricing considerations. The LumosEdge 3.0 roadmap reflects months of strategic planning by Lumos's senior leadership and engineering teams and represents a comprehensive blueprint for the company's competitive strategy over the next several years.

27. Access to the LumosEdge 3.0 product roadmap was strictly restricted to senior leadership — including the CEO, CTO, and VP of Engineering — and a small number of select senior engineers who were directly involved in the planning process and whose input was necessary to develop the technical specifications. Defendant Anand was one of the senior engineers who was granted access to the LumosEdge 3.0 roadmap by virtue of his role as Principal Engineer — ML Infrastructure and his direct involvement in the architectural planning process for the next-generation signal-generation engine. The roadmap document was stored in a restricted-access folder within Lumos's internal document-management system, and access to the folder was limited to the approximately twelve individuals who were authorized to view it.

28. Disclosure of the LumosEdge 3.0 product roadmap to a competitor would provide direct and detailed insight into Lumos's strategic direction, planned technology investments, and go-to-market timelines. A competitor armed with this information could preempt Lumos's planned product launches by racing to develop and release competing features before Lumos brings LumosEdge 3.0 to market. Such preemption would undermine Lumos's competitive position, erode its ability to differentiate its product, and potentially result in the loss of existing and prospective customers — outcomes that would cause severe and irreparable competitive harm.

Defendants' Roles, Access, and Departures

Rajiv Anand.

29. Defendant Rajiv Anand, age 37, served as Principal Engineer — ML Infrastructure at Lumos from March 2019 through September 30, 2022 — a period of approximately three and a half years. Anand was one of the most senior individual contributors in the engineering organization and was a key architect of the LumosEdge signal-generation engine. By virtue of his seniority and his central role in the design and implementation of Lumos's core ML systems, Anand had access to Lumos's most sensitive proprietary codebases, including the complete ensemble model architecture and all associated neural-network topology definitions and hyperparameter configuration files. He also had access to the model-training infrastructure, portions of the proprietary training dataset, and — as described above — the confidential LumosEdge 3.0 product roadmap. Anand's access to Lumos's proprietary technology was among the broadest of any engineer at the company.

30. In connection with the commencement of his employment at Lumos in March 2019, Anand executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above — specifically, the assignment-of-inventions clause, the three-year post-termination confidentiality and non-disclosure provision, the twelve-month non-solicitation-of-employees clause, the perpetual covenant not to use or disclose Lumos trade secrets, and the obligation to return all company materials and certify their return upon termination. A true and correct copy of Anand's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit A.

31. Anand resigned from Lumos on September 30, 2022, effective immediately, without providing the standard two-week notice period customary in the technology industry and expected by Lumos. He provided no explanation for his abrupt departure and did not disclose any plans regarding future employment or business ventures. Just three days later, on October 3, 2022, NovaSpark Technologies Inc. was incorporated in the State of Delaware under Delaware File Number 7294183, with Anand listed as CEO and co-founder. At no point did Anand return company materials to Lumos or confirm in writing the deletion of Lumos proprietary information from his personal devices, accounts, or storage media, as required by the terms of his PIIAA. Lumos sent a written demand to Anand on November 28, 2022, requesting compliance with his PIIAA return-of-materials obligations, but Anand did not respond to this demand.

Serena Mallick.

32. Defendant Serena Mallick, age 34, served as Senior Software Engineer — Data Pipelines at Lumos from June 2020 through October 14, 2022 — a period of approximately two years and four months. As described in Paragraphs 19 through 22 above, Mallick was the primary engineer responsible for building and maintaining the proprietary DataStream ETL pipeline system. She had complete access to the entire DataStream codebase, including all source code, data-schema definitions, configuration files, vendor-specific integration modules for all 43 alternative-data vendors, test suites, and internal documentation. Mallick's role gave her uniquely comprehensive knowledge of the DataStream system and the proprietary engineering insights it embodies.

33. In connection with the commencement of her employment at Lumos in June 2020, Mallick executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above. A true and correct copy of Mallick's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit B.

34. Mallick resigned from Lumos on October 14, 2022. Based on information obtained during the course of this litigation, Mallick commenced employment at NovaSpark Technologies Inc. as Vice President of Engineering just three days later, on October 17, 2022. Mallick did not disclose her plans to join NovaSpark at the time of her resignation, nor did she return company materials or confirm in writing the deletion of Lumos proprietary information from her personal devices, accounts, or storage media, as required by her PIIAA. Lumos sent a written demand to Mallick on November 28, 2022, requesting compliance with her PIIAA return-of-materials obligations, but Mallick did not respond to this demand.

Derek Yun.

35. Defendant Derek Yun, age 31, served as Machine Learning Engineer II at Lumos from January 2021 through October 21, 2022 — a period of approximately one year and ten months. As described in Paragraphs 15 through 18 and 25 above, Yun had direct access to Lumos's proprietary training datasets, model hyperparameters, and the feature-engineering libraries containing 2,847 engineered features. His daily work involved implementing and testing features, running model-training experiments, and working with the model-training pipeline — all of which required hands-on access to some of Lumos's most sensitive proprietary assets.

36. In connection with the commencement of his employment at Lumos in January 2021, Yun executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above. A true and correct copy of Yun's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit C.

37. Yun resigned from Lumos on October 21, 2022, and based on information obtained during the course of this litigation, commenced employment at NovaSpark Technologies Inc. as a founding engineer just three days later, on October 24, 2022. Like Anand and Mallick, Yun did not disclose his plans to join NovaSpark at the time of his resignation, nor did he return company materials or confirm in writing the deletion of Lumos proprietary information from his personal devices, accounts, or storage media, as required by his PIIAA. Lumos sent a written demand to Yun on November 28, 2022, requesting compliance with his PIIAA return-of-materials obligations, but Yun did not respond to this demand.

NovaSpark Technologies Inc.

38. NovaSpark Technologies Inc. was incorporated in the State of Delaware on October 3, 2022 — precisely three days after Defendant Anand's resignation from Lumos — with Defendant Anand serving as Chief Executive Officer and co-founder. According to publicly available corporate filings and NovaSpark's own website, the company is headquartered at 88 Bluxome Street, Floor 4, San Francisco, California 94107. The timing of NovaSpark's incorporation — just three days after Anand's abrupt, unannounced resignation — strongly suggests that Anand had been planning the formation of NovaSpark while still employed at Lumos and while still in possession of Lumos's most sensitive proprietary information.

39. According to NovaSpark's website and publicly available marketing materials, NovaSpark purports to offer AI-driven alternative-data analytics for hedge funds and proprietary trading desks. This description of NovaSpark's business directly overlaps with Lumos's core offering and with the specific technology at issue in this case. NovaSpark is targeting the same category of institutional clients, offering the same type of predictive analytics, and leveraging the same types of alternative data that are at the heart of the LumosEdge platform and the trade secrets that the individual Defendants exfiltrated.

40. Based on publicly available information, NovaSpark received $12 million in seed funding from Blackpine Capital Partners LP, a venture capital firm whose Managing Partner is Franklin Osei. This seed round closed on approximately November 15, 2022 — only approximately six weeks after the incorporation of NovaSpark and only weeks after the departures of all three individual Defendants from Lumos. The speed with which NovaSpark secured substantial venture funding raises serious questions about what representations were made to investors regarding NovaSpark's technology capabilities and development timeline — capabilities and a timeline that would be dramatically accelerated by the possession of Lumos's stolen trade secrets.

IT Security Investigation and Exfiltration Findings

Discovery of Anomalous Activity

41. On December 12, 2022, members of Lumos's IT security team, operating under my direct oversight, identified anomalous data-exfiltration activity during a routine review of the company's access and transfer logs. The anomalous activity dated to the final weeks of Defendants Anand's, Mallick's, and Yun's respective tenures at Lumos — specifically, the period from approximately September 26, 2022, through October 18, 2022. Upon identification of these anomalies, I directed the IT security team to conduct a comprehensive forensic investigation into the data-access and data-transfer activities of all three individual Defendants during the final 90 days of their respective employment at Lumos. The findings of that investigation are set forth below.

42. The anomalous activity was identified through Lumos's comprehensive logging and monitoring infrastructure. As described in Paragraph 8 above, Lumos maintains extensive logging of all data-access events, file transfers, data exports, cloud-upload events, USB-device connections, and email forwarding of attachments. These logs are generated by an integrated suite of security tools, including a data-loss prevention ("DLP") system that monitors and logs all outbound data transfers, an endpoint detection and response ("EDR") system deployed on all company-issued devices that logs USB device connections and local file operations, a cloud access security broker ("CASB") that monitors all interactions between company endpoints and cloud storage services, and an email security gateway that logs all outbound email attachments. Each of these systems generates timestamped, detailed log entries that identify the user, the specific files or data accessed or transferred, the destination of any transfer, and other relevant metadata. The investigation team reviewed and correlated logs from all of these systems to reconstruct the exfiltration activities described below.

Exfiltration by Rajiv Anand

43. The IT security investigation revealed that on September 26, 2022 — four days before his resignation from Lumos — Defendant Anand initiated a bulk file transfer from Lumos's internal Aldermere Software repositories to a personal Hollowell Health IT account. The transfer, which was logged by Lumos's DLP system, comprised approximately 3.2 gigabytes of files. Analysis of the DLP logs identified the transferred files as including model architecture specification documents, neural-network topology definition files, and hyperparameter configuration files for the LumosEdge signal-generation engine. These files were transferred from directories within the restricted-access ML infrastructure repository that contained the core intellectual property of the LumosEdge ensemble model architecture described in Paragraphs 11 through 14 above.

44. Additionally, on September 28, 2022 — two days before his resignation — Anand accessed the confidential "LumosEdge 3.0 Product Roadmap" document, a 37-page PDF stored in the restricted-access strategic planning folder within Lumos's internal document-management system. Lumos's email security gateway logs confirm that on the same date, Anand forwarded this 37-page PDF as an email attachment to his personal email address: r.anand.personal@vaultmail.com. The email was sent from Anand's Lumos corporate email account, and the attachment was confirmed by the security gateway logs to match the file hash of the LumosEdge 3.0 Product Roadmap document.

45. Both of the foregoing transfers — the 3.2 GB transfer to personal Hollowell Health IT on September 26, 2022, and the email forwarding of the roadmap document on September 28, 2022 — were captured and logged by Lumos's data-loss prevention and email monitoring systems. Both transfers were to personal, non-Lumos accounts that Anand controlled, and neither transfer was authorized by Lumos management. I have confirmed that no legitimate business purpose existed for Anand to transfer these materials to personal accounts, and there was no Lumos policy or practice that would have authorized such transfers. To the contrary, Lumos's information-security policies expressly prohibit the transfer of proprietary materials to personal accounts or devices.

46. Detailed analysis of the DLP logs for the 3.2 GB transfer to Westvale Technologies that the transferred files originated from directories within the ML infrastructure repository that contained the following categories of proprietary materials: (a) ensemble model architecture specifications, including the documents defining the structure and design of the LumosEdge ensemble framework; (b) neural-network topology definitions, including the specific layer configurations, activation functions, and interconnection patterns of the constituent neural-network models; and (c) hyperparameter configuration files, including the precise parameter values used to train and deploy the production LumosEdge models. These materials form the technical core of the LumosEdge system and are among Lumos's most valuable and closely guarded trade secrets.

Exfiltration by Serena Mallick

47. On October 10, 2022 — four days before her resignation from Lumos — Defendant Mallick connected a personal external USB storage device to her Lumos-issued laptop computer and cloned the entire DataStream ETL pipeline repository to that device. The connection of the USB device and the subsequent file transfer were logged by Lumos's endpoint detection and response ("EDR") system, which recorded the USB device's serial number as WD-4837291 (a Glenrock Industries portable hard drive). The EDR logs confirm that the total volume of data transferred to the USB device was approximately 780 megabytes, which is consistent with the total size of the complete DataStream repository.

48. The cloned repository contained the complete DataStream codebase, including all source code, data-schema definitions, vendor-specific parsing and transformation logic for all 43 alternative-data vendor integrations, configuration files, test suites, and associated internal documentation. As described in Paragraphs 19 through 22 above, this codebase represents the entirety of Lumos's proprietary data-ingestion and transformation infrastructure and embodies years of specialized engineering effort. Mallick had no authorization to copy the DataStream repository — or any portion of it — to a personal external storage device. No legitimate business purpose for such a transfer existed, and the transfer was in direct violation of Lumos's information-security policies, which expressly prohibit the copying of proprietary materials to personal external storage devices.

Exfiltration by Derek Yun

49. On October 18, 2022 — three days before his resignation from Lumos — Defendant Yun downloaded 47 proprietary Jupyter notebooks from Lumos's internal code repository to a personal Larkspur Technologies account. Jupyter notebooks are interactive computational documents commonly used in data science and machine learning that contain both executable code and explanatory text. The download was logged by Lumos's cloud access security broker ("CASB"), which recorded the upload of 47 files from Yun's Lumos-issued workstation to the personal Larkspur Technologies account associated with Yun's personal email address.

50. Analysis of the CASB logs and a review of the file names and directory paths of the 47 downloaded Jupyter notebooks confirmed that the notebooks contained feature-engineering code, including the mathematical formulations and software implementations for a significant portion of the 2,847 proprietary engineered features described in Paragraphs 15 through 18 above. In addition, the notebooks contained model-training scripts that reference and utilize Lumos's proprietary training dataset and hyperparameter configurations. The notebooks thus contained a concentrated distillation of some of Lumos's most valuable proprietary technology — the feature-engineering library and the model-training methodology.

51. Yun's download of these 47 notebooks to a personal cloud storage account was logged by the CASB system with full timestamps, file identifiers, and destination account details. No legitimate business justification existed for downloading these notebooks to a personal Larkspur Technologies account. The transfer was unauthorized by Lumos management and was in direct violation of Lumos's information-security policies. There was no Lumos project, assignment, or initiative that would have required Yun to transfer proprietary Jupyter notebooks to a personal account.

Conclusion of IT Security Investigation

52. In summary, the IT security investigation conducted under my oversight revealed a clear and troubling pattern: all three individual Defendants — Anand, Mallick, and Yun — exfiltrated Lumos proprietary materials to personal, non-Lumos accounts or storage devices in the days immediately preceding their respective resignations from the company. The exfiltration events occurred within a compressed timeframe — between September 26 and October 18, 2022 — and each Defendant departed Lumos within days of his or her respective exfiltration activity. All three Defendants promptly joined NovaSpark Technologies Inc., a competing startup incorporated by Anand just three days after his own resignation. The timing, coordination, and nature of these exfiltration events strongly suggest a concerted and premeditated effort by the individual Defendants to appropriate Lumos's trade secrets for use at NovaSpark. Moreover, none of the three Defendants returned company materials or confirmed the deletion of proprietary information upon departure, as required by their PIIAAs, and none responded to Lumos's written demands for compliance with these obligations.

Irreparable Harm and Competitive Impact

53. Based on my technical expertise and my intimate familiarity with the proprietary technology at issue, it is my professional opinion that the exfiltrated materials described above would enable NovaSpark Technologies Inc. to replicate core components of the LumosEdge platform in a fraction of the time and at a fraction of the cost that Lumos invested in developing those components. Specifically, I estimate that access to the ensemble model architecture, feature-engineering libraries, DataStream ETL pipeline codebase, and associated training data and documentation could save a competitor approximately 14 to 18 months of development time compared to building equivalent technology from scratch. This estimate is based on my knowledge of the complexity of the exfiltrated technology, the size and capability of the engineering teams that would be required to develop comparable systems independently, and the typical development timelines for comparable machine-learning and data-infrastructure projects in the fintech industry.

54. As described in Paragraph 39, NovaSpark's stated business — AI-driven alternative-data analytics for hedge funds and proprietary trading desks — directly competes with Lumos's core market. NovaSpark is targeting the same types of institutional clients, offering the same category of predictive-analytics services, and leveraging the same types of alternative data that are at the heart of the LumosEdge platform. The exfiltrated technology is not merely tangentially related to NovaSpark's business — it is precisely the technology that NovaSpark would need to build a competing platform.

55. Once trade secrets are disclosed to or used by a competitor, the resulting damage cannot be undone. The confidential nature of the information is permanently and irreversibly compromised. Even if the trade secrets are subsequently enjoined from further use, the knowledge gained from reviewing and analyzing the misappropriated materials cannot be "unlearned" by the individuals who have been exposed to them. Monetary damages alone would be wholly insufficient to compensate Lumos for this loss, because the value of trade secrets is inherently tied to their secrecy, and once that secrecy is lost, no amount of money can restore it. Moreover, the full extent of the competitive harm caused by the misappropriation of trade secrets is inherently difficult to quantify, as it may manifest in lost customers, lost revenue, reduced market share, impaired competitive positioning, and diminished enterprise value — consequences that are speculative and extremely difficult to measure with precision.

56. The competitive harm to Lumos is further compounded by the fact that the exfiltrated materials include not only existing technology but also the confidential LumosEdge 3.0 product roadmap. Possession of this 37-page strategic document gives NovaSpark advance knowledge of Lumos's unreleased product plans, including planned features, technical specifications, competitive positioning, and go-to-market timelines for the next several years. This forward-looking information is, if anything, even more competitively damaging than the existing technology, because it enables NovaSpark to anticipate and preempt Lumos's strategic moves before Lumos can execute on them.

57. Absent immediate injunctive relief from this Court, Lumos faces ongoing and escalating irreparable harm. Each day that passes without a restraining order is a day during which NovaSpark and the individual Defendants may continue to possess, review, analyze, and use the misappropriated materials to develop competing products, refine their technology, and solicit Lumos's existing and prospective customers. The risk of harm increases with every passing day, as NovaSpark progresses further in its development efforts and the misappropriated technology becomes more deeply embedded in NovaSpark's products and operations. Lumos's ability to obtain meaningful relief diminishes correspondingly.

Conclusion and Request for Relief

58. Based on the foregoing, Lumos Analytics Inc. respectfully requests that this Court issue a Temporary Restraining Order and Order to Show Cause Re: Preliminary Injunction requiring Defendants NovaSpark Technologies Inc., Rajiv Anand, Serena Mallick, and Derek Yun, and each of their agents, servants, employees, and all persons acting in concert or participation with them, to:

(a) Immediately cease and desist from any and all use, disclosure, dissemination, copying, or exploitation of Lumos proprietary information and trade secrets, including but not limited to the LumosEdge ensemble ML model architecture, proprietary feature-engineering libraries, DataStream ETL pipeline codebase, curated and labeled training datasets, and the LumosEdge 3.0 product roadmap;

(b) Immediately preserve and refrain from destroying, deleting, altering, modifying, overwriting, or otherwise tampering with all electronic devices, storage media, cloud accounts, email accounts, and any and all files, documents, and data that may contain or relate to Lumos proprietary information and trade secrets, including but not limited to the personal Hollowell Health IT account, the personal email account (r.anand.personal@vaultmail.com), the external USB drive bearing serial number WD-4837291, and the personal Larkspur Technologies account identified in the IT security investigation;

(c) Immediately return to Lumos all company materials in their possession, custody, or control, including all documents, files, code, data, devices, and any copies thereof, in compliance with their obligations under their respective PIIAAs; and

(d) Refrain from soliciting, recruiting, or inducing any Lumos employee to leave the company, in compliance with the non-solicitation provisions of their respective PIIAAs.

59. I declare under penalty of perjury under the laws of the State of California that the foregoing is true and correct.

Executed on January 19, 2023, at San Jose, California.

________________________________________

Rachel Fong Vice President of Engineering Lumos Analytics Inc. 4700 Orchard Parkway, Suite 300 San Jose, California 95134

Date: January 19, 2023

LUMOS ANALYTICS INC., a Delaware corporation,	Case No. 23-CV-004871
Plaintiff,	DECLARATION OF RACHEL FONG

IN SUPPORT OF PLAINTIFF'S

TEMPORARY RESTRAINING ORDER

NOVASPARK TECHNOLOGIES INC., a Delaware corporation; RAJIV ANAND, an individual; SERENA MALLICK, an individual; and DEREK YUN, an individual,	AND ORDER TO SHOW CAUSE RE: PRELIMINARY INJUNCTION
Defendants.	Date: January 19, 2023

SUPERIOR COURT OF THE STATE OF CALIFORNIA
COUNTY OF SANTA CLARA
I, Rachel Fong, declare as follows:
Introduction and Personal Background
1.  My name is Rachel Fong. I am over the age of eighteen years and am competent to testify to the facts set forth in this declaration. I have personal knowledge of each of the matters stated herein and, if called as a witness, could and would competently testify thereto.
2.  I currently serve as Vice President of Engineering at Lumos Analytics Inc. ("Lumos" or "Lumos AI"), a Delaware-incorporated, California-headquartered software-as-a-service company specializing in artificial-intelligence-driven financial analytics for institutional asset managers. Lumos's principal place of business is located at 4700 Orchard Parkway, Suite 300, San Jose, California 95134. In my capacity as Vice President of Engineering, I am responsible for the strategic direction, development, and maintenance of all of Lumos's proprietary technology platforms, including the flagship LumosEdge product described in detail below.
3.  I have over eighteen years of professional experience in software engineering, machine-learning systems, and data-infrastructure design. Prior to joining Lumos, I held senior engineering leadership positions at two prominent enterprise software companies, where I led teams responsible for building large-scale data-processing and predictive-analytics platforms. At Lumos, I oversee the engineering organization, which is the company's largest department. As of January 2023, Lumos employs approximately 340 people across all functions, and the engineering department comprises approximately 165 of those employees, including machine-learning engineers, data scientists, software engineers, data-pipeline engineers, infrastructure engineers, and quality-assurance engineers. I report directly to Lumos's Chief Executive Officer, Priya Chandrasekaran.
4.  I joined Lumos in April 2018, approximately one year after the company's founding, and have served continuously in roles of increasing responsibility since that time. Over the course of nearly five years with the company, I have developed deep and comprehensive familiarity with every aspect of Lumos's proprietary technology, including the specific trade secrets and proprietary materials at issue in this matter. During my tenure, I have been directly involved in the hiring, management, and supervision of all engineering personnel, including Defendants Rajiv Anand, Serena Mallick, and Derek Yun. I had direct familiarity with the work performed by each of these individuals during their respective tenures at Lumos, including the specific proprietary systems, codebases, datasets, and confidential business information to which they had access.
5.  I make this declaration for the purpose of providing the Court with detailed technical information regarding the proprietary technology and trade secrets at issue in this matter, and to describe the findings of the internal IT security investigation conducted by the Lumos information-security team under my oversight, which identified the specific data-exfiltration events carried out by the individual Defendants in connection with their departures from the company. The facts set forth herein are based on my personal knowledge, my review of Lumos's business records maintained in the ordinary course of business, and information provided to me by members of the Lumos IT security team who report to me.
Lumos's Business and Proprietary Technology
6.  Lumos Analytics Inc. was founded in approximately January 2017 by Priya Chandrasekaran and two co-founders with the mission of applying advanced machine-learning techniques to financial data to generate actionable, predictive analytics for institutional asset managers. Since its founding, Lumos has built and continuously refined its flagship software-as-a-service platform, branded "LumosEdge." The LumosEdge platform uses proprietary ensemble machine-learning models to ingest and analyze vast quantities of traditional financial data and alternative data, and to generate predictive signals that institutional clients — including hedge funds, mutual fund complexes, pension funds, and sovereign wealth funds — use for equity and fixed-income portfolio optimization. As of the end of fiscal year 2022, Lumos achieved annual recurring revenue ("ARR") of approximately $67 million, reflecting strong and accelerating adoption of the LumosEdge platform across the institutional asset-management industry. To date, Lumos has raised approximately $185 million in total venture capital funding through its Series D round of financing. Key funding milestones include a $40 million Series B round led by Ridgeline Ventures, which closed in August 2019, and a $72 million Series D round led by Northstar Growth Equity, which closed in June 2022. These investments reflect the substantial market value placed on Lumos's proprietary technology and competitive position.
7.  Lumos's competitive advantage in the marketplace rests fundamentally on its proprietary technology — specifically, its machine-learning model architectures, feature-engineering libraries, data-ingestion and transformation pipelines, curated training datasets, and confidential strategic product plans. This technology has been developed over the course of approximately six years at enormous expense. The cumulative investment in the technology at issue represents thousands upon thousands of engineering hours and tens of millions of dollars in direct research-and-development expenditures, not including the substantial indirect costs of recruiting and retaining the specialized engineering talent necessary to build and maintain these systems. The technology is not publicly available, has never been published or open-sourced, and is closely guarded as among the company's most valuable assets.
8.  Lumos maintains rigorous and multi-layered security protocols specifically designed to protect its proprietary technology and trade secrets from unauthorized access, use, and disclosure. These protocols include, but are not limited to: (a) role-based access controls ("RBAC") that restrict access to specific repositories, systems, and datasets based on an employee's job function and need-to-know; (b) mandatory multi-factor authentication ("MFA") for all access to internal systems, including code repositories, data warehouses, and cloud infrastructure; (c) encryption of all proprietary data and source code both at rest and in transit, using industry-standard AES-256 encryption; (d) network segmentation that isolates sensitive development and data environments from general corporate networks; (e) comprehensive logging and monitoring of all data-access events, file transfers, email attachments, and cloud-upload events through integrated data-loss prevention ("DLP"), endpoint detection and response ("EDR"), and cloud access security broker ("CASB") systems; and (f) requiring all employees who are granted access to proprietary technology or confidential business information to execute a Proprietary Information and Inventions Assignment Agreement ("PIIAA") as a condition of their employment.
9.  The PIIAA executed by each Lumos engineering employee contains several critical provisions designed to protect the company's intellectual property. Specifically, each PIIAA includes: (a) an assignment-of-inventions clause, under which the employee assigns to Lumos all right, title, and interest in any invention, improvement, development, discovery, or trade secret conceived or developed during the employee's tenure and within the scope of the employee's employment or using Lumos resources; (b) a confidentiality and non-disclosure provision that prohibits the employee from disclosing, publishing, or otherwise making available to any third party any Lumos confidential or proprietary information, which provision expressly survives termination of employment for a period of three years; (c) a non-solicitation-of-employees clause that prohibits the employee from directly or indirectly soliciting, recruiting, or inducing any Lumos employee to leave the company for a period of twelve months following the employee's separation; and (d) a covenant not to use or disclose Lumos trade secrets at any time, whether during or after employment. The PIIAA further requires the employee to return all company materials — including all documents, files, devices, code, data, and any copies thereof — upon termination of employment for any reason, and to certify in writing that all company materials have been returned and that no copies have been retained. These restrictive covenant provisions are critical to protecting Lumos's intellectual property and maintaining the confidentiality of its trade secrets.
10.  In my capacity as Vice President of Engineering, I was personally and directly involved in overseeing the development and ongoing maintenance of each of the five categories of proprietary technology described in the following paragraphs. I am intimately familiar with the technical details, competitive significance, and protective measures associated with each category.
Category 1: LumosEdge Ensemble ML Model Architecture
11.  The LumosEdge ensemble machine-learning model architecture is the core technological engine of the LumosEdge platform. This architecture consists of a proprietary neural-network topology that combines multiple distinct model types — including deep neural networks, gradient-boosted decision trees, and proprietary attention-based transformer variants — into an ensemble framework specifically designed and optimized for predictive signal generation in equity and fixed-income portfolio contexts. The architecture incorporates proprietary hyperparameter configurations that have been carefully tuned through extensive experimentation and validation. The specific design choices embodied in the ensemble architecture — including the number and types of constituent models, the topology of each model, the methods by which model outputs are combined and weighted, the proprietary loss functions employed, and the precise hyperparameter values — collectively represent Lumos's most valuable and closely guarded technical asset. The ensemble architecture is what enables LumosEdge to generate the predictive signals that differentiate Lumos's offering from those of its competitors.
12.  The LumosEdge ensemble model architecture was developed over a multi-year period, beginning in approximately mid-2018 and continuing through the present, by a core team of senior machine-learning engineers working under my supervision. Defendant Rajiv Anand was a central member of this core team. Anand served as Principal Engineer — ML Infrastructure at Lumos from March 2019 through September 30, 2022, and was one of the key architects of the LumosEdge signal-generation engine. In that role, Anand had deep, hands-on access to the complete ensemble model architecture, including all neural-network topology definitions, hyperparameter configuration files, model-training pipelines, and associated technical documentation. His contributions to the architecture were significant, but they were made as a Lumos employee, using Lumos resources, within the scope of his employment, and subject to the PIIAA he executed upon hire.
13.  The ensemble model architecture, including all topology definitions and hyperparameter configurations, is not disclosed publicly in any form. It is not described in any published papers, blog posts, patent applications, or marketing materials. The architecture is stored exclusively on access-restricted internal code repositories that require both role-based authorization and multi-factor authentication to access. Access to the specific repositories containing the model architecture specifications is limited to a small group of senior ML engineers and engineering leadership. The architecture represents significant competitive value — based on my experience and knowledge of the relevant field, I estimate that recreating equivalent models of comparable performance from scratch, without access to the proprietary architecture, would require a well-resourced team of experienced ML engineers a minimum of two to three years of intensive research and development, at a cost of many millions of dollars.
14.  The loss of confidentiality in the LumosEdge ensemble model architecture would allow a competitor to replicate the core signal-generation capabilities of the LumosEdge platform without incurring the same development time and cost that Lumos invested. Such a competitor would be able to stand up a functionally equivalent predictive-analytics engine in a dramatically compressed timeframe, enabling it to enter Lumos's core market and compete directly for Lumos's institutional clients on the basis of technology that Lumos developed at great expense.
Category 2: Proprietary Feature-Engineering Libraries
15.  Lumos has developed and maintains a proprietary set of feature-engineering libraries containing 2,847 individually engineered features specifically designed for financial signal generation. In machine learning, "features" are the specific, mathematically defined transformations applied to raw input data to create the structured inputs that are fed into predictive models. Lumos's 2,847 proprietary features transform raw financial data — including price, volume, order-book, fundamental, macroeconomic, and sentiment data — as well as alternative data from non-traditional sources, into precisely calibrated inputs for the LumosEdge ensemble models. Each feature embodies a specific hypothesis about informative patterns in financial data, and each has been individually researched, designed, implemented, tested, and validated by Lumos's data science and engineering teams.
16.  The feature-engineering library is the product of years of extensive research, experimentation, and iterative testing. The process of developing effective features for financial signal generation is one of the most time-intensive and intellectually demanding aspects of building predictive ML models for financial applications. It requires deep domain expertise in both quantitative finance and machine learning, combined with access to high-quality data and substantial computational resources. The specific features in Lumos's library, their mathematical formulations, the methods by which they are combined and selected for use in different model configurations, and the proprietary insights they embody about the structure of financial data, are all proprietary and confidential. The feature library is maintained in access-restricted internal repositories and is not disclosed publicly in any form.
17.  Defendant Derek Yun served as Machine Learning Engineer II at Lumos from January 2021 through October 21, 2022. In that role, Yun had direct access to the proprietary feature-engineering libraries as part of his daily work. Yun's responsibilities included implementing new features, testing feature performance, and integrating features into the model-training pipeline. Through this work, Yun gained intimate knowledge of the feature library's contents, including the mathematical formulations, code implementations, and performance characteristics of a significant portion of the 2,847 engineered features.
18.  Access to the complete set of 2,847 proprietary features would provide a competitor with an enormous head start in developing a competing predictive-analytics platform. Rather than investing the years of research and experimentation necessary to identify and validate effective features, a competitor in possession of Lumos's feature library could immediately deploy those features in its own models, capturing the value of Lumos's investment without bearing the cost. The feature-engineering process is, in my professional judgment, one of the most valuable and difficult-to-replicate components of any financial machine-learning system, and Lumos's library represents an asset of extraordinary competitive significance.
Category 3: DataStream ETL Pipeline Codebase
19.  Lumos has developed a proprietary data-ingestion and transformation system known internally as "DataStream." DataStream is an extract, transform, and load ("ETL") pipeline consisting of a proprietary codebase and accompanying data-schema definitions that manage the end-to-end process of ingesting, cleaning, normalizing, enriching, and delivering data from 43 distinct alternative-data vendors into the LumosEdge platform. Alternative data — including satellite imagery analytics, web-scraping feeds, credit-card transaction aggregates, geolocation data, social-media sentiment data, and other non-traditional data sources — is a critical input to the LumosEdge models, and the DataStream pipeline is the system that makes this data usable.
20.  The DataStream pipeline was built and maintained by a specialized data-pipelines engineering team within the Lumos engineering organization. Defendant Serena Mallick served as Senior Software Engineer — Data Pipelines at Lumos from June 2020 through October 14, 2022. During her tenure, Mallick was the primary engineer responsible for building and maintaining the DataStream system. She designed and implemented many of the vendor-specific integration modules, wrote core components of the data-transformation logic, and maintained the data-schema definitions that govern how data flows through the pipeline. Mallick had complete and unrestricted access to the entire DataStream codebase, all associated schema definitions, configuration files, and technical documentation.
21.  The DataStream codebase includes proprietary logic for parsing, deduplicating, reconciling, and enriching data from each of the 43 vendor integrations. Each vendor integration represents significant engineering effort, as each vendor delivers data in its own format, with its own quirks, quality issues, delivery schedules, and API specifications. The DataStream codebase embodies proprietary knowledge about vendor-specific data quality issues — including specific fields that are unreliable, known data gaps, latency characteristics, and edge cases in vendor data formats — as well as the optimal transformation methodologies that Lumos developed through extensive testing to address these issues. The integrations with these 43 vendors took years of cumulative engineering effort to develop, test, validate, and maintain, and the resulting codebase represents a formidable barrier to entry for any competitor seeking to build a comparable data-ingestion capability.
22.  The complete DataStream repository, including all source code, configuration files, data-schema definitions, test suites, and internal documentation, represents approximately 780 megabytes of data. This codebase is stored in a private, access-restricted Git repository hosted on Lumos's internal Aldermere Software instance. Access to the DataStream repository is limited to members of the data-pipelines team and senior engineering leadership, and all access is logged and monitored.
Category 4: Curated and Labeled Training Dataset
23.  Lumos maintains a proprietary training dataset of approximately 14.6 terabytes of curated and labeled historical financial and alternative data. This dataset was assembled over the entire course of Lumos's operations, beginning in 2017, and represents the cumulative output of years of data acquisition, curation, cleaning, labeling, and annotation by Lumos's data science team. The dataset includes historical market data, fundamental data, macroeconomic indicators, and alternative data from dozens of sources, all of which have been processed, aligned, and structured into a format optimized for training the LumosEdge ensemble models.
24.  Critically, the training dataset includes not only raw data that has been licensed from third-party data providers under agreements that contain strict restrictions on redistribution and sublicensing, but also substantial quantities of proprietary labels, annotations, and derived data that were created by Lumos's data science team through extensive manual and semi-automated processes. These proprietary labels and annotations — which encode expert judgments about market regimes, event classifications, signal quality, and other domain-specific concepts — are among the most valuable components of the dataset and are entirely the product of Lumos's own intellectual effort. The training dataset is essential for training and validating the LumosEdge models, and its quality and comprehensiveness are a significant competitive differentiator. The investment in data acquisition alone — including licensing fees paid to third-party data vendors over the past six years — amounts to tens of millions of dollars, and the investment in curation and labeling represents thousands of hours of specialized data-science work.
25.  Defendants Anand and Yun each had access to portions of this training dataset through their respective work on the ML infrastructure and model-training pipelines. Anand, as Principal Engineer — ML Infrastructure, had broad access to the training-data infrastructure, including the systems used to store, version, and serve training data to model-training jobs. Yun, as Machine Learning Engineer II, had access to specific subsets of the training data used in the model-training experiments he conducted as part of his daily work. Both individuals had sufficient access to obtain significant portions of the proprietary training dataset, including the proprietary labels and annotations.
Category 5: Confidential Product Roadmap — LumosEdge 3.0
26.  In addition to its existing proprietary technology, Lumos maintains confidential product roadmap documents that detail the company's strategic product direction for the period 2023 through 2025. These documents include detailed specifications for the as-yet-unreleased "LumosEdge 3.0" product version, which represents the next major evolution of the LumosEdge platform. The roadmap document is a 37-page PDF that contains, among other things: planned new features and capabilities, detailed technical specifications for new model architectures and data-processing systems, competitive analysis of Lumos's market position relative to identified competitors, go-to-market timelines and launch strategies, target customer segments, and pricing considerations. The LumosEdge 3.0 roadmap reflects months of strategic planning by Lumos's senior leadership and engineering teams and represents a comprehensive blueprint for the company's competitive strategy over the next several years.
27.  Access to the LumosEdge 3.0 product roadmap was strictly restricted to senior leadership — including the CEO, CTO, and VP of Engineering — and a small number of select senior engineers who were directly involved in the planning process and whose input was necessary to develop the technical specifications. Defendant Anand was one of the senior engineers who was granted access to the LumosEdge 3.0 roadmap by virtue of his role as Principal Engineer — ML Infrastructure and his direct involvement in the architectural planning process for the next-generation signal-generation engine. The roadmap document was stored in a restricted-access folder within Lumos's internal document-management system, and access to the folder was limited to the approximately twelve individuals who were authorized to view it.
28.  Disclosure of the LumosEdge 3.0 product roadmap to a competitor would provide direct and detailed insight into Lumos's strategic direction, planned technology investments, and go-to-market timelines. A competitor armed with this information could preempt Lumos's planned product launches by racing to develop and release competing features before Lumos brings LumosEdge 3.0 to market. Such preemption would undermine Lumos's competitive position, erode its ability to differentiate its product, and potentially result in the loss of existing and prospective customers — outcomes that would cause severe and irreparable competitive harm.
Defendants' Roles, Access, and Departures
Rajiv Anand.  
29.  Defendant Rajiv Anand, age 37, served as Principal Engineer — ML Infrastructure at Lumos from March 2019 through September 30, 2022 — a period of approximately three and a half years. Anand was one of the most senior individual contributors in the engineering organization and was a key architect of the LumosEdge signal-generation engine. By virtue of his seniority and his central role in the design and implementation of Lumos's core ML systems, Anand had access to Lumos's most sensitive proprietary codebases, including the complete ensemble model architecture and all associated neural-network topology definitions and hyperparameter configuration files. He also had access to the model-training infrastructure, portions of the proprietary training dataset, and — as described above — the confidential LumosEdge 3.0 product roadmap. Anand's access to Lumos's proprietary technology was among the broadest of any engineer at the company.
30.  In connection with the commencement of his employment at Lumos in March 2019, Anand executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above — specifically, the assignment-of-inventions clause, the three-year post-termination confidentiality and non-disclosure provision, the twelve-month non-solicitation-of-employees clause, the perpetual covenant not to use or disclose Lumos trade secrets, and the obligation to return all company materials and certify their return upon termination. A true and correct copy of Anand's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit A.
31.  Anand resigned from Lumos on September 30, 2022, effective immediately, without providing the standard two-week notice period customary in the technology industry and expected by Lumos. He provided no explanation for his abrupt departure and did not disclose any plans regarding future employment or business ventures. Just three days later, on October 3, 2022, NovaSpark Technologies Inc. was incorporated in the State of Delaware under Delaware File Number 7294183, with Anand listed as CEO and co-founder. At no point did Anand return company materials to Lumos or confirm in writing the deletion of Lumos proprietary information from his personal devices, accounts, or storage media, as required by the terms of his PIIAA. Lumos sent a written demand to Anand on November 28, 2022, requesting compliance with his PIIAA return-of-materials obligations, but Anand did not respond to this demand.
Serena Mallick.  
32.  Defendant Serena Mallick, age 34, served as Senior Software Engineer — Data Pipelines at Lumos from June 2020 through October 14, 2022 — a period of approximately two years and four months. As described in Paragraphs 19 through 22 above, Mallick was the primary engineer responsible for building and maintaining the proprietary DataStream ETL pipeline system. She had complete access to the entire DataStream codebase, including all source code, data-schema definitions, configuration files, vendor-specific integration modules for all 43 alternative-data vendors, test suites, and internal documentation. Mallick's role gave her uniquely comprehensive knowledge of the DataStream system and the proprietary engineering insights it embodies.
33.  In connection with the commencement of her employment at Lumos in June 2020, Mallick executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above. A true and correct copy of Mallick's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit B.
34.  Mallick resigned from Lumos on October 14, 2022. Based on information obtained during the course of this litigation, Mallick commenced employment at NovaSpark Technologies Inc. as Vice President of Engineering just three days later, on October 17, 2022. Mallick did not disclose her plans to join NovaSpark at the time of her resignation, nor did she return company materials or confirm in writing the deletion of Lumos proprietary information from her personal devices, accounts, or storage media, as required by her PIIAA. Lumos sent a written demand to Mallick on November 28, 2022, requesting compliance with her PIIAA return-of-materials obligations, but Mallick did not respond to this demand.
Derek Yun.  
35.  Defendant Derek Yun, age 31, served as Machine Learning Engineer II at Lumos from January 2021 through October 21, 2022 — a period of approximately one year and ten months. As described in Paragraphs 15 through 18 and 25 above, Yun had direct access to Lumos's proprietary training datasets, model hyperparameters, and the feature-engineering libraries containing 2,847 engineered features. His daily work involved implementing and testing features, running model-training experiments, and working with the model-training pipeline — all of which required hands-on access to some of Lumos's most sensitive proprietary assets.
36.  In connection with the commencement of his employment at Lumos in January 2021, Yun executed the company's standard PIIAA, including all of the restrictive covenant provisions described in Paragraph 9 above. A true and correct copy of Yun's executed PIIAA is attached to the concurrently filed Declaration of Katherine Everett as Exhibit C.
37.  Yun resigned from Lumos on October 21, 2022, and based on information obtained during the course of this litigation, commenced employment at NovaSpark Technologies Inc. as a founding engineer just three days later, on October 24, 2022. Like Anand and Mallick, Yun did not disclose his plans to join NovaSpark at the time of his resignation, nor did he return company materials or confirm in writing the deletion of Lumos proprietary information from his personal devices, accounts, or storage media, as required by his PIIAA. Lumos sent a written demand to Yun on November 28, 2022, requesting compliance with his PIIAA return-of-materials obligations, but Yun did not respond to this demand.
NovaSpark Technologies Inc.
38.  NovaSpark Technologies Inc. was incorporated in the State of Delaware on October 3, 2022 — precisely three days after Defendant Anand's resignation from Lumos — with Defendant Anand serving as Chief Executive Officer and co-founder. According to publicly available corporate filings and NovaSpark's own website, the company is headquartered at 88 Bluxome Street, Floor 4, San Francisco, California 94107. The timing of NovaSpark's incorporation — just three days after Anand's abrupt, unannounced resignation — strongly suggests that Anand had been planning the formation of NovaSpark while still employed at Lumos and while still in possession of Lumos's most sensitive proprietary information.
39.  According to NovaSpark's website and publicly available marketing materials, NovaSpark purports to offer AI-driven alternative-data analytics for hedge funds and proprietary trading desks. This description of NovaSpark's business directly overlaps with Lumos's core offering and with the specific technology at issue in this case. NovaSpark is targeting the same category of institutional clients, offering the same type of predictive analytics, and leveraging the same types of alternative data that are at the heart of the LumosEdge platform and the trade secrets that the individual Defendants exfiltrated.
40.  Based on publicly available information, NovaSpark received $12 million in seed funding from Blackpine Capital Partners LP, a venture capital firm whose Managing Partner is Franklin Osei. This seed round closed on approximately November 15, 2022 — only approximately six weeks after the incorporation of NovaSpark and only weeks after the departures of all three individual Defendants from Lumos. The speed with which NovaSpark secured substantial venture funding raises serious questions about what representations were made to investors regarding NovaSpark's technology capabilities and development timeline — capabilities and a timeline that would be dramatically accelerated by the possession of Lumos's stolen trade secrets.
IT Security Investigation and Exfiltration Findings
Discovery of Anomalous Activity
41.  On December 12, 2022, members of Lumos's IT security team, operating under my direct oversight, identified anomalous data-exfiltration activity during a routine review of the company's access and transfer logs. The anomalous activity dated to the final weeks of Defendants Anand's, Mallick's, and Yun's respective tenures at Lumos — specifically, the period from approximately September 26, 2022, through October 18, 2022. Upon identification of these anomalies, I directed the IT security team to conduct a comprehensive forensic investigation into the data-access and data-transfer activities of all three individual Defendants during the final 90 days of their respective employment at Lumos. The findings of that investigation are set forth below.
42.  The anomalous activity was identified through Lumos's comprehensive logging and monitoring infrastructure. As described in Paragraph 8 above, Lumos maintains extensive logging of all data-access events, file transfers, data exports, cloud-upload events, USB-device connections, and email forwarding of attachments. These logs are generated by an integrated suite of security tools, including a data-loss prevention ("DLP") system that monitors and logs all outbound data transfers, an endpoint detection and response ("EDR") system deployed on all company-issued devices that logs USB device connections and local file operations, a cloud access security broker ("CASB") that monitors all interactions between company endpoints and cloud storage services, and an email security gateway that logs all outbound email attachments. Each of these systems generates timestamped, detailed log entries that identify the user, the specific files or data accessed or transferred, the destination of any transfer, and other relevant metadata. The investigation team reviewed and correlated logs from all of these systems to reconstruct the exfiltration activities described below.
Exfiltration by Rajiv Anand
43.  The IT security investigation revealed that on September 26, 2022 — four days before his resignation from Lumos — Defendant Anand initiated a bulk file transfer from Lumos's internal Aldermere Software repositories to a personal Hollowell Health IT account. The transfer, which was logged by Lumos's DLP system, comprised approximately 3.2 gigabytes of files. Analysis of the DLP logs identified the transferred files as including model architecture specification documents, neural-network topology definition files, and hyperparameter configuration files for the LumosEdge signal-generation engine. These files were transferred from directories within the restricted-access ML infrastructure repository that contained the core intellectual property of the LumosEdge ensemble model architecture described in Paragraphs 11 through 14 above.
44.  Additionally, on September 28, 2022 — two days before his resignation — Anand accessed the confidential "LumosEdge 3.0 Product Roadmap" document, a 37-page PDF stored in the restricted-access strategic planning folder within Lumos's internal document-management system. Lumos's email security gateway logs confirm that on the same date, Anand forwarded this 37-page PDF as an email attachment to his personal email address: r.anand.personal@vaultmail.com. The email was sent from Anand's Lumos corporate email account, and the attachment was confirmed by the security gateway logs to match the file hash of the LumosEdge 3.0 Product Roadmap document.
45.  Both of the foregoing transfers — the 3.2 GB transfer to personal Hollowell Health IT on September 26, 2022, and the email forwarding of the roadmap document on September 28, 2022 — were captured and logged by Lumos's data-loss prevention and email monitoring systems. Both transfers were to personal, non-Lumos accounts that Anand controlled, and neither transfer was authorized by Lumos management. I have confirmed that no legitimate business purpose existed for Anand to transfer these materials to personal accounts, and there was no Lumos policy or practice that would have authorized such transfers. To the contrary, Lumos's information-security policies expressly prohibit the transfer of proprietary materials to personal accounts or devices.
46.  Detailed analysis of the DLP logs for the 3.2 GB transfer to Westvale Technologies that the transferred files originated from directories within the ML infrastructure repository that contained the following categories of proprietary materials: (a) ensemble model architecture specifications, including the documents defining the structure and design of the LumosEdge ensemble framework; (b) neural-network topology definitions, including the specific layer configurations, activation functions, and interconnection patterns of the constituent neural-network models; and (c) hyperparameter configuration files, including the precise parameter values used to train and deploy the production LumosEdge models. These materials form the technical core of the LumosEdge system and are among Lumos's most valuable and closely guarded trade secrets.
Exfiltration by Serena Mallick
47.  On October 10, 2022 — four days before her resignation from Lumos — Defendant Mallick connected a personal external USB storage device to her Lumos-issued laptop computer and cloned the entire DataStream ETL pipeline repository to that device. The connection of the USB device and the subsequent file transfer were logged by Lumos's endpoint detection and response ("EDR") system, which recorded the USB device's serial number as WD-4837291 (a Glenrock Industries portable hard drive). The EDR logs confirm that the total volume of data transferred to the USB device was approximately 780 megabytes, which is consistent with the total size of the complete DataStream repository.
48.  The cloned repository contained the complete DataStream codebase, including all source code, data-schema definitions, vendor-specific parsing and transformation logic for all 43 alternative-data vendor integrations, configuration files, test suites, and associated internal documentation. As described in Paragraphs 19 through 22 above, this codebase represents the entirety of Lumos's proprietary data-ingestion and transformation infrastructure and embodies years of specialized engineering effort. Mallick had no authorization to copy the DataStream repository — or any portion of it — to a personal external storage device. No legitimate business purpose for such a transfer existed, and the transfer was in direct violation of Lumos's information-security policies, which expressly prohibit the copying of proprietary materials to personal external storage devices.
Exfiltration by Derek Yun
49.  On October 18, 2022 — three days before his resignation from Lumos — Defendant Yun downloaded 47 proprietary Jupyter notebooks from Lumos's internal code repository to a personal Larkspur Technologies account. Jupyter notebooks are interactive computational documents commonly used in data science and machine learning that contain both executable code and explanatory text. The download was logged by Lumos's cloud access security broker ("CASB"), which recorded the upload of 47 files from Yun's Lumos-issued workstation to the personal Larkspur Technologies account associated with Yun's personal email address.
50.  Analysis of the CASB logs and a review of the file names and directory paths of the 47 downloaded Jupyter notebooks confirmed that the notebooks contained feature-engineering code, including the mathematical formulations and software implementations for a significant portion of the 2,847 proprietary engineered features described in Paragraphs 15 through 18 above. In addition, the notebooks contained model-training scripts that reference and utilize Lumos's proprietary training dataset and hyperparameter configurations. The notebooks thus contained a concentrated distillation of some of Lumos's most valuable proprietary technology — the feature-engineering library and the model-training methodology.
51.  Yun's download of these 47 notebooks to a personal cloud storage account was logged by the CASB system with full timestamps, file identifiers, and destination account details. No legitimate business justification existed for downloading these notebooks to a personal Larkspur Technologies account. The transfer was unauthorized by Lumos management and was in direct violation of Lumos's information-security policies. There was no Lumos project, assignment, or initiative that would have required Yun to transfer proprietary Jupyter notebooks to a personal account.
Conclusion of IT Security Investigation
52.  In summary, the IT security investigation conducted under my oversight revealed a clear and troubling pattern: all three individual Defendants — Anand, Mallick, and Yun — exfiltrated Lumos proprietary materials to personal, non-Lumos accounts or storage devices in the days immediately preceding their respective resignations from the company. The exfiltration events occurred within a compressed timeframe — between September 26 and October 18, 2022 — and each Defendant departed Lumos within days of his or her respective exfiltration activity. All three Defendants promptly joined NovaSpark Technologies Inc., a competing startup incorporated by Anand just three days after his own resignation. The timing, coordination, and nature of these exfiltration events strongly suggest a concerted and premeditated effort by the individual Defendants to appropriate Lumos's trade secrets for use at NovaSpark. Moreover, none of the three Defendants returned company materials or confirmed the deletion of proprietary information upon departure, as required by their PIIAAs, and none responded to Lumos's written demands for compliance with these obligations.
Irreparable Harm and Competitive Impact
53.  Based on my technical expertise and my intimate familiarity with the proprietary technology at issue, it is my professional opinion that the exfiltrated materials described above would enable NovaSpark Technologies Inc. to replicate core components of the LumosEdge platform in a fraction of the time and at a fraction of the cost that Lumos invested in developing those components. Specifically, I estimate that access to the ensemble model architecture, feature-engineering libraries, DataStream ETL pipeline codebase, and associated training data and documentation could save a competitor approximately 14 to 18 months of development time compared to building equivalent technology from scratch. This estimate is based on my knowledge of the complexity of the exfiltrated technology, the size and capability of the engineering teams that would be required to develop comparable systems independently, and the typical development timelines for comparable machine-learning and data-infrastructure projects in the fintech industry.
54.  As described in Paragraph 39, NovaSpark's stated business — AI-driven alternative-data analytics for hedge funds and proprietary trading desks — directly competes with Lumos's core market. NovaSpark is targeting the same types of institutional clients, offering the same category of predictive-analytics services, and leveraging the same types of alternative data that are at the heart of the LumosEdge platform. The exfiltrated technology is not merely tangentially related to NovaSpark's business — it is precisely the technology that NovaSpark would need to build a competing platform.
55.  Once trade secrets are disclosed to or used by a competitor, the resulting damage cannot be undone. The confidential nature of the information is permanently and irreversibly compromised. Even if the trade secrets are subsequently enjoined from further use, the knowledge gained from reviewing and analyzing the misappropriated materials cannot be "unlearned" by the individuals who have been exposed to them. Monetary damages alone would be wholly insufficient to compensate Lumos for this loss, because the value of trade secrets is inherently tied to their secrecy, and once that secrecy is lost, no amount of money can restore it. Moreover, the full extent of the competitive harm caused by the misappropriation of trade secrets is inherently difficult to quantify, as it may manifest in lost customers, lost revenue, reduced market share, impaired competitive positioning, and diminished enterprise value — consequences that are speculative and extremely difficult to measure with precision.
56.  The competitive harm to Lumos is further compounded by the fact that the exfiltrated materials include not only existing technology but also the confidential LumosEdge 3.0 product roadmap. Possession of this 37-page strategic document gives NovaSpark advance knowledge of Lumos's unreleased product plans, including planned features, technical specifications, competitive positioning, and go-to-market timelines for the next several years. This forward-looking information is, if anything, even more competitively damaging than the existing technology, because it enables NovaSpark to anticipate and preempt Lumos's strategic moves before Lumos can execute on them.
57.  Absent immediate injunctive relief from this Court, Lumos faces ongoing and escalating irreparable harm. Each day that passes without a restraining order is a day during which NovaSpark and the individual Defendants may continue to possess, review, analyze, and use the misappropriated materials to develop competing products, refine their technology, and solicit Lumos's existing and prospective customers. The risk of harm increases with every passing day, as NovaSpark progresses further in its development efforts and the misappropriated technology becomes more deeply embedded in NovaSpark's products and operations. Lumos's ability to obtain meaningful relief diminishes correspondingly.
Conclusion and Request for Relief
58.  Based on the foregoing, Lumos Analytics Inc. respectfully requests that this Court issue a Temporary Restraining Order and Order to Show Cause Re: Preliminary Injunction requiring Defendants NovaSpark Technologies Inc., Rajiv Anand, Serena Mallick, and Derek Yun, and each of their agents, servants, employees, and all persons acting in concert or participation with them, to:
(a) Immediately cease and desist from any and all use, disclosure, dissemination, copying, or exploitation of Lumos proprietary information and trade secrets, including but not limited to the LumosEdge ensemble ML model architecture, proprietary feature-engineering libraries, DataStream ETL pipeline codebase, curated and labeled training datasets, and the LumosEdge 3.0 product roadmap;
(b) Immediately preserve and refrain from destroying, deleting, altering, modifying, overwriting, or otherwise tampering with all electronic devices, storage media, cloud accounts, email accounts, and any and all files, documents, and data that may contain or relate to Lumos proprietary information and trade secrets, including but not limited to the personal Hollowell Health IT account, the personal email account (r.anand.personal@vaultmail.com), the external USB drive bearing serial number WD-4837291, and the personal Larkspur Technologies account identified in the IT security investigation;
(c) Immediately return to Lumos all company materials in their possession, custody, or control, including all documents, files, code, data, devices, and any copies thereof, in compliance with their obligations under their respective PIIAAs; and
(d) Refrain from soliciting, recruiting, or inducing any Lumos employee to leave the company, in compliance with the non-solicitation provisions of their respective PIIAAs.
59.  I declare under penalty of perjury under the laws of the State of California that the foregoing is true and correct.
Executed on January 19, 2023, at San Jose, California.
________________________________________
Rachel Fong Vice President of Engineering Lumos Analytics Inc. 4700 Orchard Parkway, Suite 300 San Jose, California 95134
Date: January 19, 2023
LUMOS ANALYTICS INC., a Delaware corporation, | Case No. 23-CV-004871
Plaintiff, | DECLARATION OF RACHEL FONG
IN SUPPORT OF PLAINTIFF'S
v. | EX PARTE APPLICATION FOR
TEMPORARY RESTRAINING ORDER
NOVASPARK TECHNOLOGIES INC., a Delaware corporation; RAJIV ANAND, an individual; SERENA MALLICK, an individual; and DEREK YUN, an individual, | AND ORDER TO SHOW CAUSE RE: PRELIMINARY INJUNCTION
Defendants. | Date: January 19, 2023

SUPERIOR COURT OF THE STATE OF CALIFORNIA

COUNTY OF SANTA CLARA

IN SUPPORT OF PLAINTIFF'S

TEMPORARY RESTRAINING ORDER

Record

Parties

Dates

Amounts