Hedge-Fund AI Labs, Research Agents, and Academic Lineage: What the Public Record Actually Connects

Source status: second-pass synthesis completed 2026-08-15, with date-scoped personnel, conference, and title-blind media refreshes through 2026-09-02 from firm-controlled pages, public paper records, university or personal academic profiles, conference and vendor/customer artifacts, and dated hiring evidence.

Source files: sources/06-industry-verticals/hedge-fund-ai-labs-agents-lineage-2026-08-15-raw.md, sources/06-industry-verticals/hedge-fund-ai-labs-academic-lineage-2026-08-14-raw.md, sources/06-industry-verticals/hedge-fund-ai-landscape-cross-audit-2026-08-12-raw.md, sources/06-industry-verticals/hedge-fund-domain-model-agentic-job-audit-2026-raw.md, sources/06-industry-verticals/hedge-fund-ai-conference-video-mining-2026-08-14-raw.md, sources/13-multimodal-sources/researcher-social-index-and-title-blind-leads-2026-08-18-raw.md, sources/13-multimodal-sources/title-blind-manager-lab-and-governance-routes-2026-09-02-raw.md, sources/13-multimodal-sources/conference-and-event-gap-expansion-2026-09-02-raw.md, sources/13-multimodal-sources/person-first-media-gap-expansion-2026-09-02-raw.md, sources/13-multimodal-sources/helix-ae-regional-ai-quant-role-pass-2026-09-02-raw.md, and research/06-industry-verticals/hedge-fund-ai-labs-agents-lineage-2026.csv.

Additional source note: sources/13-multimodal-sources/varsity-tech-louis-liu-chinese-podcast-agent-pass-2026-09-02-raw.md and sources/13-multimodal-sources/regional-quant-ai-followup-nayt-blocktech-nex-hillsdale-2026-09-02-raw.md.

Chart data: chart-ready cohort data. The external evidence links below are the citable public record.

Executive Summary

• Five named or lab-like structures are publicly identifiable: Bridgewater AIA Labs, CFM ML Lab, Millennium AI Lab, Schonfeld FE AI Lab, and XTX XTY Labs. Balyasny, Man Group, Two Sigma, Citadel, and BlackRock expose adjacent central groups or platforms without the same naming convention.
• The public research-agent functions divide into four lanes: information filtering and retrieval; hypothesis generation and coding; forecasting and event monitoring; and workflow automation around documents, earnings, spreadsheets, and research operations.
• Academic lineage is visible for a subset of named people. MIT, Stanford, Brown, Johns Hopkins, Columbia, NYU, TTIC, Berkeley, CMU, and ENS appear in the public records. Doctoral advisors are public for some researchers, including Rohan Alur, Gene Li, Lei Chen, Jianbo Chen, and Honglin Yuan.
• Published work reveals methods, not firm deployment. AIA Forecaster, BloombergGPT, CFM’s financial NER case study, graph-learning papers, and RLVR work are public artifacts. None proves that a paper’s method is a live trading signal.
• The missing data is itself consistent: public sources rarely connect a named researcher to a production model, a model to a portfolio workflow, and that workflow to independently audited results.

What counts as a lab

The word “lab” is not used consistently across the industry. This draft uses three categories. A named lab is a firm-controlled source that calls the unit a lab. A lab-like group is a central AI or ML organization with a research or platform remit but no stable lab name. A no named lab found result means that the reviewed public sources exposed roles, infrastructure, or AI strategy without a named unit.

That distinction matters because a job description can reveal intended architecture while saying nothing about whether the role was filled. A researcher’s publication can reveal methods while saying nothing about current firm assignment. A vendor case study can describe a customer workflow while leaving the internal model registry and permission map private.

Named labs and their public remit

September 2 title-blind expansion: four distinct public structures

The Hildene episode page names CLOver as Hildene’s proprietary investment-analytics platform and lists the personnel involved in a structured-credit video. Hildene’s current team page separately lists Mark Rubidge as Vice President of Quantitative Research & AI. The pages establish a named analytics platform and a current role; they do not connect Rubidge to CLOver or disclose architecture, data, permissions, or outcomes. The YouTube player remains a recording-recovery target.

BlackRock’s AI Labs page identifies Rachel Schutt and Stephen Boyd as co-heads, names academic advisers, and describes applications of statistics, machine learning, optimization, stochastic control, and decision theory across retirement, trading, alternatives, and ETFs. Its public catalogue includes portfolio construction, hyper-parameter optimization, performance attribution, market design, and open-source optimization work. This is a first-party research-remit and academic-network disclosure; it does not connect each artifact to a live portfolio or publish a complete model and permission inventory.

QRT’s QRT Labs page describes a 2026 partnership with Imperial, Cambridge, and Oxford, with more than 70 doctoral and postdoctoral researchers in the initial phase, three university-based centres, seminars, and an annual conference. The stated themes include mathematical foundations of AI, agentic systems, complex decisions, high-performance computing, cybersecurity, hardware, and advanced modelling. This is evidence of research infrastructure and talent development, not a disclosure of a trading model or capital authority.

PIMCO’s UK Stewardship Code report describes proprietary and customized AI tools, employee training, a showcase of 25 AI solutions attended by more than 1,000 staff, an AI governance group, and an AI controls group reviewing customized solutions. It explicitly frames AI as support for human expertise and names data quality, hallucination, security, privacy, recordkeeping, and MNPI as control concerns. The August 2026 Accrued Interest episode adds a current first-party transcript route, but not a model disclosure. PIMCO is retained here as an institutional asset-management governance comparator, not as hedge-fund evidence.

September 2 Singapore expansion: Varsity Tech and a public agent loop

The Varsity Tech website describes a Singapore AI-native quantitative trading and investment-technology firm whose platform combines strategy research, backtesting, AI research agents, and automated execution. The site says the platform currently targets U.S. equities and is expanding to U.S. futures and multi-asset systematic trading. This is a first-party product description; it does not establish a verified hedge-fund vehicle, customer access to execution, model inventory, data rights, or audited results.

The public Louis Liu profile and Luma event listing identify Liu as a Varsity Tech cofounder and advertise a bilingual July 1, 2026 session on an end-to-end factor workflow: hypothesis generation, code, backtest, keep/kill, and iteration, with discussion of agent architecture and prompting. The event listing is marked past and does not expose a recording or slides. These pages establish a named personnel/event route, not the delivery or technical truth of the advertised workflow.

The Chinese-language APL-XLink podcast page adds a 52-minute public transcript route. It presents Louis L. as a Varsity Tech cofounder / Quandora founding-team member and describes product iteration, family-office and small-fund conversations, a lean team, and a preference for experienced engineers who use AI critically rather than relying on unreviewed code generation. The page also contains founder-reported financing and historical-trading assertions; those are not promoted because the page does not supply independent confirmation, risk-adjusted evaluation, or a verified account. The company’s social post adds an OpenClaw Singapore FinTech Edition / Moomoo SG partner route and says Liu discussed agentic quantitative research and execution. It remains social and event evidence, not a model or permission disclosure.

This cohort is useful precisely because the public artifact chain is different from a named research lab: product page → founder profile → event listing → Chinese podcast transcript → partner/event post. It exposes the research loop and staffing philosophy, but leaves legal entity, customer, model, data, execution authority, and outcome questions open. See the full source note.

September 2 regional follow-up: strategy and hiring surfaces without a named lab

The Nayt careers page describes a Singapore proprietary trading firm with HFT market making, cross-exchange funding arbitrage, systematic mid-frequency long/short, DeFi, and an in-house toolset spanning risk, databases, strategy control, execution, and analytics. Its public personnel route labels Jos Pol Co-Founder & Head of Trading, while a LiquidityTech post advertises his guest appearance on market edge and microstructure. The latter is an episode-discovery lead; the recording and transcript were not resolved in this pass. Public headcount bands differ between the careers page and LinkedIn, so the discrepancy remains explicit.

The BlockTech senior quantitative-researcher listing describes an end-to-end research path from hypotheses and dataset construction through model training, backtesting, deployment, monitoring, and iteration, including price-prediction, signal, execution, and anomaly-detection models. The company history adds the Amsterdam and Singapore operating footprint and a firm-reported 100-person milestone. These are public role and company statements; no model, data-vendor, permission, or performance chain is disclosed.

The NEX Capital site exposes a Quant AI Analyst role that mentions LLMs, reinforcement learning, deep neural networks, unconventional datasets, signal generation, portfolio optimization, and backtesting/live trading infrastructure. The firm’s legal vehicle, named team, filled role, model registry, and audited results remain unresolved, so it is tracked as an investment-firm watchlist lead rather than a verified hedge-fund lab.

The Canadian Hillsdale systematic-equities role routes ML and LLM adoption through quantitative research, portfolio construction, risk modeling, attribution, constraints, and human review. Hillsdale’s research library lists published work on ML stock selection, quantitative asset management, and LLM risk. A public Xavier Mootoo profile connects a current quantitative-equity-research role to York/Krembil deep learning and time-series research, while his personal page provides additional publication context. The public chain supports personnel and research-lineage mapping; it does not show that the academic or open-source work is a Hillsdale production model. See the regional follow-up source note.

The public record therefore contains several different organizational designs: a named research-and-investment lab, a scientific ML lab connected to academia, an experimentation environment, an investment-team training program, a research division with foundation-model hiring, and a small founder-led platform that advertises an agent-to-backtest loop. Those are organizational descriptions, not a capability league table.

September 2 Bridgewater and CFM personnel refresh

The Bridgewater/CFM route note adds date-scoped evidence to the lab roster and organizational map. Bridgewater’s Nina Lozinski profile identifies her as Co-Head of AI & ML Investment Strategy and reports the firm’s account that AIA Labs grew from six people in 2023 to more than 50 scientists, engineers, and investors. A current Greg Jensen profile connects his Managing CIO remit to AIA Labs, while a separate first-party article identifies Suri Bandler as an Architect on the Technology Team. CFM’s October 2025 publication describes the ML Lab as part of its research teams and as a bridge between academic AI/ML work and firm research.

The Bridgewater headcount and titles are firm-reported and date-scoped. The CFM statement is a CFM-hosted publication description. They do not reveal a complete roster, reporting lines, model inventory, training corpus, data permissions, deployment stage, or investment attribution, and they should not be interpreted as a comparative judgment.

BlackRock Systematic: additional primary and title-blind evidence

The BlackRock Quantitative Investing careers page adds a Systematic-specific operating and hiring surface. It describes 200+ professionals across San Francisco, New York, and London; teams spanning systematic active equity and fixed income; machine learning, AI, large data, alternative data, peer review, investment attribution, and responsible AI use; and academic partnerships. Its Systematic Active Equity description specifically names trading activity, internet searches, economic forecasts, and demographic trends as inputs. This is recruiting and firm-process evidence, not proof that every described capability is live or that open roles were filled.

BlackRock’s macro machine-learning paper names Raffaele Savi, Stephanie Lee, Phil Green, Thomas Logan, Ronald Kahn, Michael Pyle, and Michael Pensky. The described techniques include pooling across macro assets and countries, regularized models, geospatial mapping, language-model-enhanced embeddings, local economic data, and learn-to-rank methods for relative-value applications. The paper exposes model-design choices for low-breadth, regime-sensitive macro problems; it does not expose model code, data rights, production permissions, or independent validation.

The CFA Institute award record and practice brief connect “Thematic Investing: A Risk-Based Perspective” to Ked Hogan, Ronald Kahn, Robert Luo, Emmanuel Candès, Trevor Hastie, and Asher Spector. Public posts by Jeff Shen and Raffaele Savi identify the Systematic and AI Labs affiliation split. The paper and award establish research lineage, not a production model or return attribution.

The Resonanz Spotlight episode with Jeff Shen is a title-blind podcast route dated July 2, 2025. Its publisher describes LLM use in investment workflows and cross-disciplinary teams spanning engineering, finance, and academia. Savi’s Client Research Summit post names Applied Intuition CEO Qasar Younis and describes client discussions, research, and demos. Both are public discussion or event evidence; neither discloses a complete model inventory, commercial integration, or autonomous authority.

Millennium’s current partnership disclosure adds a second, separate layer to its lab evidence. The August 2026 Anthropic announcement describes forward-deployed Anthropic engineers working with Millennium technology and risk teams on a supervised digital risk analyst. The stated workflow covers explaining daily risk changes, interrogating data, retaining context across interactions, and surfacing risk insights across asset classes; Millennium also says it will test Anthropic frontier models against firm work. This is partnership and evaluation evidence, not proof of autonomous risk-limit changes, order submission, or performance attribution.

The title-blind personnel layer is also broader than “AI researcher.” Millennium publicly names an Equity AI engineer working with LLMs on fundamental-equity portfolio-manager processes, an AI engineer building agentic tools for investment professionals, a fixed-income/commodities engineer building data-discoverability tools and chatbots, an infrastructure leader with AI-platform responsibility, and dated ML-research, quantitative-modeling, and Data & AI conference routes. The detailed capture note is here.

September 2 regional expansion: AI-native language without a verified lab roster

The Helix AI Capital website adds a Singapore/Hong Kong proprietary-trading surface, not a verified hedge-fund record. It describes systematic equities, crypto derivatives, and global futures, and claims deep-learning models for equity statistical arbitrage, reinforcement learning for crypto execution and market making, and transformer sequence models plus cross-asset graph networks for futures. It also claims tick-level and alternative data, petabyte-scale training data, in-house data warehouses, and GPU training clusters. These are specific first-party model and infrastructure statements, but the public record does not independently verify the operating entity, production deployment, data rights, capital, or results.

Helix’s careers page lists roles titled AI Scientist — Foundation Models for Markets, Machine Learning Researcher — Alpha Models, and LLM Scientist — Research Automation, alongside quantitative, data-infrastructure, and execution-engineering roles. The LinkedIn company page showed a 2–10 employee range and two visible employees at capture, while the firm’s site says fewer than thirty people. The discrepancy is retained as a verification flag. No named AI owner or academic lineage is inferred. The separate crypto quant researcher posting and portfolio-manager posting describe on-chain data, order books, funding, open interest, market microstructure, realistic backtests, capacity, and production readiness; both were closed when captured. They show hiring intent, not a filled role or live strategy.

The AE Capital careers page adds an Australian title-blind control case. It describes machine learning and advanced data science for fully automated systematic strategies. Its separate Junior Engineer / Quantitative Developer / Quantitative Analyst posting identifies global FX and futures, real-time market-data systems, model testing and monitoring, production support, C#/.NET, SQL, AWS, Python, and a progression from engineering into quantitative development and analysis. The posting was closed at capture. Neither page names a model, data vendor, filled employee, or audited outcome. The useful signal is the workflow vocabulary: AI work can be found through data, execution, monitoring, and production titles even when the title itself does not contain “AI.” See the regional role-pass note.

Models, embeddings, and agent artifacts that must not be collapsed

The lab-and-lineage view is only one slice of the evidence. A reconciliation against the earlier model and agent audit adds a second layer: named models, embeddings, public research papers, agent workflows, and runtime infrastructure. These artifacts are not interchangeable. A paper can show a method; a recording can show what a presenter said; a job listing can show intended architecture; and a firm page can describe bounded use. None automatically proves a live signal or autonomous capital authority.

Bridgewater PAT and the Interrupt 2026 recording

Yes, PAT belongs in this article, but the earlier version did not carry over the additional source: LangChain’s Interrupt 2026 recording library and the Bridgewater PAT session. Bridgewater’s official PAT page dates the recording to May 19, 2026 and identifies Brendan McManus, Michael Ran, and Santi Weight as presenters.

The recording adds implementation detail to the short written lab page: PAT is described as an investigation-oriented research agent rather than a trading engine; a chat layer gathers context, searches unstructured documents and time series, asks clarifying questions, and creates a plan; a coding layer converts that plan into Python/Pandas work; parallel sub-agents generate code; a dependency graph controls execution; static analysis and validation agents check results; and cached execution reduces repeated computation. The presenters also describe entitlement-aware context and tools, millions of documents, tens of millions of time-series records, human-audited benchmarks, and a pull-request loop for improving agent behavior.

Those details are high-value public evidence about architecture and controls, but they remain a firm-presented account. The recording does not publish the code, benchmark denominator, model versions, permissions matrix, production error rate, or independent performance attribution.

Bridgewater AIA Labs: primary audio and date-scoped personnel

The later Greg Jensen interview, now locally captured from the canonical Acast audio, adds a separate operating account. Jensen describes a Secure Garden in which research knowledge must be written, translated into algorithms, and represented in human- and computer-readable form (43:29–45:16). He says AIA was created as an AI-centered “idea factory” with different talent, security separation, and independence from Pure Alpha’s existing process (45:50–48:12). He also describes a post-COVID reorganization intended to distribute decision-making more broadly (68:45–69:38). These are executive statements, not an independent audit of permissions or returns.

The personnel timeline needs a date boundary. Bridgewater’s Greg Jensen profile identifies him as managing CIO for the Alpha Engine and AIA Labs. Jasjeet Sekhon’s profile now identifies his former Bridgewater Chief Scientist/Head of AI role and current Google DeepMind position. Older source records that list Sekhon as current Bridgewater staff should therefore be read as historical or date-scoped. Bridgewater’s public pages do not expose a complete current AIA Labs roster.

Bridgewater’s current Nina Lozinski profile provides a stronger date-scoped personnel surface: it identifies her as Co-Head of AI & ML Investment Strategy, a Partner, and a contributor to building AIA Labs. The page says she jointly leads investment-research decisions within AIA Labs and records her University of Chicago computational-and-applied-mathematics background. This resolves current public role evidence, not a complete team roster or an independent assessment of the lab’s investment impact.

A title-blind personnel pass recovered a 2019 MIT IDSS seminar whose description identifies Jasjeet Sekhon as Head of Causal Inference at Bridgewater at the time. Sekhon distinguishes predictive ML from causal-effect estimation and discusses heterogeneous treatment effects, high-dimensional covariates, and meta-learners (09:20–09:35; 20:42–21:01; 28:03–28:14). The route is useful for academic lineage and methods mapping. It is date-scoped and does not establish current employment, Bridgewater model ownership, portfolio use, permissions, or performance. See the capture note.

The earlier 2016 CODE plenary recording provides a separate academic route for Sekhon’s work on heterogeneous treatment effects, outcome-versus-effect modeling, adaptive experimentation, and validation (01:51–11:31; 35:57–40:33; 47:05–48:42). The recording does not establish Bridgewater employment at the time, Bridgewater use of the methods, or reliable speaker ownership for every later discussion segment. It is retained as date-scoped methodology and lineage context only. See the capture note.

The Sackler Big Data Colloquium recording is another academic route, focused on large-scale randomized experiments, heterogeneous effects, researcher degrees of freedom, blocking, adaptive experimentation, and validation (00:00–14:40; 39:03–48:42). It does not establish Bridgewater employment at the time, Bridgewater use of the methods, or a firm model. See the capture note.

Reconciled artifact register

The official-channel audit adds a separate G-Research media layer: a captioned NeurIPS 2022 video linking market prediction, ML conferences, and recruiting; an Alex Davies symposium recording with a firm-written academic-lineage cross-check; and a dated Michael I. Jordan lecture on AI, uncertainty, data, and finance. These strengthen the conference and academic-network map, not the case for a named G-Research investment model or deployment.

This correction changes the article’s inventory but not its evidentiary rule: the safe unit is “public artifact plus stated function plus evidence boundary.” It is not “firm has AI,” “firm has a finance GPT,” or “firm has autonomous trading.”

September 2 title-blind and person-first additions

The conference expansion note adds distinct event surfaces for CFM, Robeco, Balyasny, Bridgewater, AQR, Marshall Wace, Systematica, Squarepoint, and QRT. The most specific additions are a Gartner session naming Balyasny’s Avanti Goenka around metadata automation and AI enablement, a Bridgewater first-party AI Today and Tomorrow page naming Jasjeet Sekhon, and a CUATS agenda that places CFM and Robeco researchers alongside a separate LLM multi-agent finance session. These establish dated public event or programme records, not firm deployment.

The person-first expansion note adds guest-name routes that ordinary hedge-fund searches miss: Jane Street’s Yaron Minsky technology talks; Numerai founder Richard Craib; CFM Chairman Jean-Philippe Bouchaud; AQR CTO Neal Pawar and former AQR personnel; and Two Sigma technology leaders Camille Fournier, Jeff Reback, and former CTO Alfred Spector. It also captures former Point72 and Bridgewater personnel routes. Current-versus-former affiliation is kept explicit, and transcript capture remains a separate queue; these routes should not be treated as current firm evidence until the role date and source content are rechecked.

The thin-firm title-blind pass adds several distinct evidence types. Voleon co-founder and CIO Jon McAuliffe’s business-podcast transcript describes a systematic machine-learning and data operation and gives a Harvard–Berkeley–D. E. Shaw lineage lead. PDT founder Peter Muller’s 2026 interview contributes founder and research-culture context, while a separate Vinesh Jha interview preserves a former-PDT route into alternative-data cleaning, point-in-time controls, and earnings-call NLP. GSA receives two separate historical conference records for Gordon Ritter’s supervised/reinforcement-learning and model-risk presentation routes. Aspect’s Ian McWilliam is identified publicly as a Senior Researcher, Machine Learning, and Man AHL’s first-party archive names Martin Luk and Slavi Marinov around an explicit ChatGPT-versus-word-counting financial-sentiment experiment. Winton gains historical academic sponsorship and researcher-lineage routes through the Alan Turing Institute and STFC. These additions expand the map of public personnel and research artifacts; they do not establish current production systems or performance.

ML researchers, colleges, advisors, and papers

The table below only includes people for whom the public record supplies a reasonably direct professional or academic trail. “Advisor” means a named doctoral advisor or principal investigator where the source states one. A school connection without an advisor is not upgraded into a PI relationship.

How to read academic lineage

Academic lineage is useful for tracing methods and collaboration networks. It is not a proxy for individual quality or firm capability. For example, graph representation learning appears in the public work of people connected to Citadel and Two Sigma, while CFM’s academic network centers on high-dimensional statistics, statistical physics, and data science. Those overlaps make certain research questions plausible. They do not show that a firm uses a particular paper in production.

The same rule applies to Bridgewater’s AIA papers. The AIA Forecaster paper describes agentic search, a supervisor agent, and statistical calibration. The 2026 expert-judgment work describes human annotations and specialized training for financial document filtering. These are direct public artifacts of the lab’s research program. They are not evidence that the disclosed systems independently allocate capital.

September 2 low-coverage firm research-process routes

The underexplored-firm route note adds methodological context without expanding the named-lab list. D. E. Shaw’s 2023 Machine Teaching publication describes optimizer-assisted portfolio reasoning, uncertainty inspection, and iterative human-machine interaction, while explicitly saying that different strategies use tailored optimization tools. Arrowstreet’s current page describes a technology team, advanced computing, portfolio optimization, and real-time signals but does not name an AI model or lab. Winton’s 2017 essay records a historical view of AI and non-stationary markets. Aspect and Dimensional add podcast and research leadership surfaces, not lab disclosures.

The D. E. Shaw publication is archived privately in checkpoint b18e34ca. These routes show why absence of a lab name should not be treated as absence of computational or research infrastructure, but they do not establish a current AI system, production authority, or investment result.

The fresh title-blind pass adds further non-lab controls. AlphaSimplex’s Form ADV Part 2A describes a highly automated proprietary-quantitative process but does not call it AI or GenAI. PanAgora’s careers archive exposes an Equity Data Science role spanning Alpha Research, Portfolio Construction, and software teams, while Rokos’s recruiting post exposes Technology and Quant hiring streams. Capula’s public jobs page displays quantitative-strategy and technology role labels. These are useful organizational and hiring surfaces, but none establishes a named lab, filled researcher, model, data rights, production permission, or investment outcome.

What the research agents do

The public record describes research agents as bounded systems around evidence, code, models, and workflow state. It does not establish a common definition of “agent,” and it does not show that an agent can place an order, change a portfolio, or bypass an investment professional.

Cross-cohort characteristics that can be charted

The useful charts are categorical and source-linked. They should not produce a maturity score or rank firms by AI quality.

1. Lab naming versus agent function

Plot firms as rows and code the lab status against the agent lane: retrieval, hypothesis generation, forecasting, or workflow automation. This shows whether the public organization is described as a research lab, an experimentation environment, a platform group, or an investment-team program.

2. Academic institution to firm network

Create a bipartite network using only public person-level links: MIT → Bridgewater, Two Sigma, Citadel, GMO; Stanford → Citadel, Two Sigma, BlackRock; Brown/Johns Hopkins → Millennium; Columbia → BlackRock; TTIC → Two Sigma; ENS/NYU → CFM. Add a second edge from each person to a named doctoral advisor where the source supplies one. This chart is a collaboration and training map, not a talent ranking.

3. Research artifact type by firm

Use a categorical bar or matrix with: firm technical report, peer-reviewed paper, model card, public code, conference talk, customer case study, job description, and firm biography. Bridgewater and CFM have multiple artifact types; Acadian and Arrowstreet currently expose more hiring and workflow architecture; Schonfeld exposes a named program with no public lab paper located in this pass.

4. Agent control surface

Code whether the public source mentions human approval, evaluation, source citation, scoped tools, default-deny execution, audit logs, cost telemetry, or deployment stage. Balyasny, Bridgewater, Millennium, Schonfeld, Acadian, and Arrowstreet expose different subsets. The chart should show controls named in sources, not infer control quality.

5. Research topic to disclosed workflow

Map paper topics to public workflow categories, with an explicit “inference” label: graph learning → network discovery or structured data; forecasting and calibration → event forecasting; NLP/LLM → document filtering and retrieval; RL and optimization → candidate strategy or execution research; formal methods → agent constraints. No edge should be labeled “alpha” unless a firm-controlled source makes that connection explicitly.

Evidence boundaries

This draft does not rank firms or assign a better-or-worse position on any dimension. It records where public evidence exists and where the public record is silent.

Current employment, current title, and entity identity require separate checks. Citadel and Citadel Securities remain separate in the ledger. Academic affiliations are not employment. A vendor-hosted customer story is attributed to the vendor and customer. A job description is hiring intent. A paper is a publication or technical artifact, not a deployment log.

The unresolved question is which firms will eventually disclose a reproducible chain from source data, to model or agent, to evaluation, to human decision, to portfolio or execution outcome. This pass found no public source that completes that chain for a named research agent.

What This Means for Your Organization

If you are comparing AI programs, ask for the same fields this ledger preserves: the named research owner, academic or practitioner lineage, task definition, source and data rights, evaluation set, tool permissions, human approval point, deployment stage, and rollback path. A public lab name is useful context. It is not a substitute for those operating details.

If this raised questions specific to your organization, the next useful step is to build the same evidence map against your own AI portfolio and research workflow.

Key Data Points

September 2 role and lab-surface refresh

The latest title-blind search adds useful distinctions between an explicit lab, an ML hiring lane, and a general quantitative-engineering route. Event Horizon Labs’ current research page describes 2026 work on agentic search, objective-function design, execution, and public market-data experiments. Its AI Research Engineer posting specifies hypothesis-forming agents, market experiments, model improvement, reward design, retained knowledge, and leaderboard/evaluation infrastructure; its infrastructure posting adds scheduler, checkpointing, versioned-data, reproducibility, and execution-monitoring responsibilities. The site names aggregate prior affiliations rather than individuals. Its blog explicitly labels its posts machine-generated, so those posts remain discovery or architecture context rather than independent research evidence.

PDT provides a different public surface: Applied ML Scientist, Quantitative Researcher, and Research Engineer roles sit beside a stated research-to-live-automated-trading process. Systematica’s firm page separates Research, Technology, and Trading and describes alternative-data and proprietary technology work, but does not name a current AI lab or model. A Jain Global AI Research Intern listing describes LLM/agent implementation, evaluation, and unstructured-data work; it is hiring evidence, not evidence of a lab, filled role, or live investment system. These records add organizational and role vocabulary without supporting an across-firm ranking or a model-capability conclusion.

September 2 regional personnel and research-process refresh

The latest independent pass adds an explicit Dymon Asia AI-intern description, an Arrowpoint ML-engineer role, and a Brevan Howard front-office data-engineer route. Dymon’s page names model evaluation, prototyping, prompting, LangChain/Hugging Face, governance, and deployment vocabulary inside fundamental-equities teams; Arrowpoint’s role covers predictive ML/deep learning, alternative-data ETL, AWS deployment, monitoring, and the model lifecycle; Brevan’s role covers data infrastructure for research, trading, and portfolio decision support. These are public hiring surfaces, not evidence of filled personnel, a named lab, or live investment authority.

The same pass adds RQI’s named executive transcript, where David Walsh describes extensive ML in research, nonlinear patterns, human-led idea generation, portfolio and risk construction, and a possible future role for agentic systems. It also adds Hadron Capital’s public platform description, which names a gradient-boosted ML alpha model, purged walk-forward validation, and executable trade generation, alongside a founder biography linking University of Alberta and CERN/ATLAS experience. Hadron’s page is self-description and RQI’s is attributed first-party commentary; neither supplies an independently reproducible model or performance record.

PanAgora’s research archive, chatbot/quantitative-investing article, and Crowell Prize announcement provide Boston-linked publication and judging routes. They show public research activity involving chatbot adoption, conference-call text, and deep learning on executive presentations, but do not establish a current AI lab, production model, or academic-advisor lineage. The full regional route inventory is in the capture note.

September 2 frontier lab and role refresh

G-Research publicly separates quantitative research and machine learning, engineering, technology innovation, and open-source teams. Grace Investment Machine uses the explicit “technology company and research lab” description and lists agentic research, signal discovery, model evaluation, backtesting, execution, risk controls, automated training, deployment, and inference work. These are new lab and role surfaces, but remain first-party organizational and hiring evidence rather than a verified roster or deployment log.

Amundi’s Fixed Income Investment Lab role adds a detailed comparator route linking rates and credit signals, alternative and unstructured data, Transformers and embeddings, bias-aware backtesting, drift monitoring, rollback, and model acceptance criteria. WorldQuant’s Singapore/Sydney portfolio-manager role links systematic strategies, broad datasets, internal research conferences, and AI/ML opportunities. One Eleven Capital and SINTRO add European quantitative-fund and systematic-trading role surfaces. None of these pages establishes a named person’s current remit, a model’s live status, or an independently measured result.

September 2 role-surface refresh: AXQ and control evidence

The role-surface capture note adds a distinct AXQ agent-harness route to the lab-and-agent inventory. The Agent Harness Research Engineer posting names prompt, tool calling, memory, workflow orchestration, evaluation, RAG, and coding-agent components, then connects them to feature generation, signal validation, backtest analysis, and research reporting. Separate AXQ postings describe financial-data lineage and production operations: the Lead Financial Data Engineer posting names temporal semantics, alternative-data quality, and support for research, backtests, and live trading, while the Production Engineer posting names model deployment support, execution and risk monitoring, incident response, runbooks, and observability.

This is a public role architecture, not a verified lab roster. It does not show which roles are filled, which models are live, who owns the agent harness, what data is licensed, or whether any system has capital authority. The same pass adds Moreton’s named ML-platform and AI-engineering personnel surface, CC&L’s NER/knowledge-graph/data-quality role, and Infinitus’s non-AI technology route. CC&L’s application page states that AI was not used in applicant screening at capture; that is a recruitment control signal and not an investment-system claim. These records should remain separate from named-lab evidence until staffing and deployment are independently resolved.

September 2 academic control pass: evaluating agentic trading

The agentic-trading evaluation note adds a research-control layer to the lab inventory. HKU Business School’s Agentic Trader report describes a common-environment live-market experiment and cautions that static reasoning benchmarks do not necessarily predict live-market behavior. The Agentic Trading survey emphasizes protocol completeness and reproducibility; the quantitative-trading survey separates factor mining, signal discovery, portfolio construction, execution, and risk; TradeLens adds trace-grounded cost and decision-value diagnostics; and AlphaCrafter provides a public Miner/Screener/Trader harness implementation.

These sources supply evaluation fields for the cross-firm map—time-consistent splits, transaction costs, execution semantics, trace retention, permission boundaries, verification, and persistent memory. They are not evidence of any named manager’s adoption or of transferable performance. The sources remain separate from the named-lab list.

September 2 emerging AI-lab and personnel surfaces

The emerging-firm and regional capture note adds several public surfaces that should be tracked separately from the larger named-manager lab inventory. Binomial Technologies’ firm page names Ilay H. Ibrahimzadeh as Portfolio Manager and Lead Engineer and describes a proprietary AI system, Lattice, for quantitative modelling, risk diagnostics, and workflow automation. The page also states that the firm is in an OpenAI pilot. These are firm-authored statements; no independent confirmation, model specification, or data-permission record was found in this pass.

KelAI’s Y Combinator profile names founder Jeremie Cohen and states former WorldQuant portfolio-management and Millennium ML-team experience. The profile describes a research-agent loop covering idea generation, code, alternative-data analysis, backtesting, validation, monitoring, and PM feedback. The profile says PMs retain portfolio ownership, but its client deployment and live-signal statements remain company claims.

Magnum Opus’ public team and infrastructure pages name Cameron Bennion, Matt Gore, and Joseph Lackland and describe in-house AI research/execution, decision logging, risk gates, and human confirmation before orders. These controls are useful fields for the lab taxonomy; they do not independently verify model operation, performance, or control effectiveness.

Nymbus’ official team page adds a Canadian personnel and lineage cluster: Gabriel Cefaloni’s AI/portfolio-management remit, Mathieu Poulin-Brière’s prior Perseus quantitative-manager role, Jessica Martins’ prior Tower Research FX role and quant-research/automation remit, Olivier Cyr-Choinière’s research role, and Jean-Philippe Lejeune’s model-development and risk-tool remit. The page does not disclose a model inventory, training corpus, or agent permission structure.

Two Japan routes widen the non-English evidence layer. Mizuho First Financial Technology publicly describes AI-enabled quantitative research using machine learning, multilayer neural networks, GPU/Hadoop, and cloud infrastructure. A historical MUFG/MTEC corporate account describes NLP, machine learning, and deep learning combined with human review before execution. Sumitomo Mitsui DS’s current quant-group page supplies a staffing and dated AUM control but does not make a specific AI/GenAI claim there. Historical, employer-reported, and role-page evidence remain distinct from a current lab roster or deployed system.

Sources

The full second-pass source ledger and prior academic-lineage ledger are maintained in the repository with this draft. The earlier firm-by-firm synthesis remains here.

The citations in the tables link to firm-controlled pages, university or personal academic profiles, public paper records, or explicitly attributed vendor/customer materials. Retrieval dates and evidence boundaries are preserved in the source ledger.

State of AI — Executive Briefings August 2026