Halluminate's 9-Person Sell to 4 Labs: The Finance-RL Pool Is 693

On October 1, 2026, Halluminate closed a $30M Series A led by Oak HC/FT with nine people on the payroll, a mid-eight-figure revenue run rate, and four of the top five closed-source US AI labs paying for its finance-specific RL environments. CEO Jerry Wu said the money goes to systems engineering and AI research hires. The problem is that the exact profile Halluminate needs, an ex-IB/PE operator who can also spec a reward function, barely exists at scale.

This article maps that pool, names where it concentrates, and shows why Halluminate's customer list and its recruiting list are the same list.

Why Halluminate's raise is a sourcing story, not a funding story

The binding constraint on vertical RL environment companies is headcount, not capital. Anthropic's leaders have discussed spending more than $1B per year on RL environments, while the US pool of people who hold a Research Engineer, Research Scientist, or Member of Technical Staff title and list Reinforcement Learning as a skill sits at 693 in Refolk's index.

An RL environment, for the uninitiated, is a programmatic sandbox where a model takes actions, gets scored against a ground-truth reward, and learns. For finance that means things like Halluminate's Westworld Due Diligence benchmark, released August 2026: 88 tasks drawn from anonymized PE transactions, written and reviewed by practicing deal professionals. The highest average frontier-model score on Westworld was 51%. That gap is why four of five frontier labs are buying.

693

The scarcity isn't the researcher. Scale AI said in February that nearly half of its new data-training projects involve RL environments, which pulled every lab, broker, and labeling vendor into the same hiring lane. The scarcity is the researcher who also understands a leveraged buyout model. That dual-pedigree atom is what a finance LLM engineer sourcing search is really hunting.

The 693-person pool, broken down

Here is the dataset recruiters should be building against, pulled from Refolk's index of professional profiles cross-referenced with public disclosures from Halluminate's Series A announcement.

Read row six twice. In a 25-profile sample of the 693-person pool, 13 people sit at OpenAI, Anthropic, Google DeepMind, or Meta. Those are also the four frontier labs most likely to be among Halluminate's four paying customers. The company that just sold to its market now has to poach from its market. Expect non-poach clauses and warrant provisions in those lab contracts to become a public story inside twelve months.

The dual pedigree problem no Boolean captures

The scarce atom isn't the RL researcher, it's the ex-Goldman VP or ex-McKinsey associate who can also write Python and spec a reward function. Halluminate's team page lists Meta, Scale AI, Capital One Labs, McKinsey, and Goldman Sachs as prior employers; that mix, not the headcount, is the moat.

Standard sourcing stacks do not surface this profile. LinkedIn Recruiter asks you to pick one skill tree. Boolean strings for ("reinforcement learning" OR "RL environments") AND ("Goldman Sachs" OR "McKinsey") return mostly false positives: investment bankers who took a weekend Coursera course, or ML engineers who did a summer analyst stint in 2014.

What actually works is describing the person the way a founder would describe them to a cofounder. Three sentences, not three filters. That is the exact gap Refolk closes: describe the dual-pedigree hire in plain English ("someone who shipped RL training code at a frontier lab or a top AI startup, and spent at least two years in investment banking M&A or private equity deal execution") and get back a ranked shortlist across GitHub, LinkedIn, and the open web.

The five sub-profiles inside the 693

Not every name in the pool is equally reachable. For an AI evals recruiter mapping Halluminate's competitive set, the pool fragments into five practical buckets:

1. Frontier-lab RL researchers with finance side quests. The rarest. Quant-adjacent researchers who still have a finance stint on the resume.
2. Ex-labeling-vendor environment builders. Sepal AI (YC S24, acquired by Mercor February 2026) and Deeptune (acquired July 2026) alumni. They have shipped eval harnesses in production and are now well into golden-handcuff territory at a $2B ARR parent.
3. Capital One Labs alumni. The only large US bank in Halluminate's own founding resume mix, which makes it the only proven feeder into a frontier-adjacent vertical-RL startup.
4. Computer use agent researcher types. The HUD, Veris AI, Plato, and Bespoke Labs cohort of environment-native startups. They know agentic eval design cold, but most have never seen a DCF.
5. Prime Intellect Environments Hub contributors. Unaffiliated open-source RL environment authors who self-identify publicly. Legitimate sourcing surface, usually underpaid.

Capital One Labs is the under-indexed feeder

Treat Capital One Labs the way sourcers treated Google Brain in 2017: small, concentrated, and about to disperse. It is the one large US bank that appears on Halluminate's own team page alongside Meta, Scale AI, McKinsey, and Goldman Sachs, which is the clearest public signal that its alumni ship into frontier-adjacent startups.

The practical move is to pull the Capital One Labs alumni list, filter for people who left in the last 24 months, and cross-reference against GitHub activity on agent eval repos. A Halluminate hiring process, or a competitor's, that doesn't start there is leaving the easiest wins on the table.

The acquired-labeling-vendor liquidity window

The first wave of poachable talent isn't at the frontier labs, it's at the labeling vendors that just got acquired. Sepal AI closed into Mercor in February 2026; Deeptune closed in July 2026. Mercor reportedly hit $2B ARR in June 2026 and now publishes the APEX agent benchmarks, which means two acquired teams are being merged into one agent-eval roadmap, a known source of regrettable attrition.

Why these cohorts matter for a Halluminate competitor, or for Halluminate itself:

• They have shipped RL environment code in production, not just research notebooks.
• They understand frontier-lab buying behavior from the vendor side.
• Acquisition liquidity resets their risk tolerance. A nine-person vertical startup looks reasonable when the acquiring parent just doubled in size around them.
• Mercor is also publicly hiring for its Deeptune RL environments team in New York as of August 2026, which means the comparable talent flow is already visible.

Halluminate matched four-of-five lab coverage with roughly one-twelfth the headcount of Surge AI.

Surge, bootstrapped at around 110 employees, quietly passed Scale in revenue. Halluminate did it with nine. The margin story is the recruiting story: frontier labs pay $85 to $200+ per expert hour on human-data contracts from vendors like Surge, Mercor, Handshake, and Turing, while vertical specialists sell packaged environments at software margins. That gap funds frontier-lab-competitive compensation for the dual-pedigree hire.

The comp benchmark Halluminate will be measured against

The ceiling comp for an RL environments engineer in 2026 is Mechanize, founded by former Epoch AI researchers and backed by Nat Friedman and Jeff Dean, which reportedly pays engineers $500,000 and sits at a $750M valuation. Halluminate will be measured against that number by every candidate it approaches.

That matters for sourcers because it compresses the negotiating range. The 693-person pool knows what Mechanize pays. Any outreach that opens with a vague "competitive compensation" line gets ignored. The first message either names a band at or above the Mechanize number, or it names an equity story that pencils out against a mid-eight-figure ARR company with four frontier-lab logos and $38.5M in total funding.

3.7%

For recruiters building Halluminate hiring pipelines or competitive maps, the practical shortlist shape looks like this:

• A working subset of the 693 pool who are not under active frontier-lab retention.
• Sepal AI and Deeptune alumni, with acquisition anniversary dates mapped.
• Capital One Labs ML research alumni from the last 24 months.
• Prime Intellect Environments Hub contributors with at least one merged environment.
• Practicing M&A bankers or PE associates who have published code on GitHub in the last year.

Everything beyond that is noise.

How to run the search without burning the pool

Three rules, learned the hard way by anyone who has recruited into a vertical this small:

1. No mass InMail. The pool is small enough that two badly-worded templates in the same month tank your reply rate for a quarter.
2. Lead with the benchmark, not the role. Westworld Due Diligence's 51% ceiling is a more compelling opener than any job description. Candidates in this pool read benchmark papers for fun.
3. Map before you message. Know which four labs Halluminate sells to before you poach from the fifth. Know which Sepal alum reports to which Mercor VP before you reach out.

Rule three is where narrow mapping tools earn their keep. Describe the exact slice ("Sepal AI alumni now at Mercor who have committed to a public RL environments repo in the last six months") and get back the actual names, not a Boolean approximation.

FAQ

How many people in the US could realistically fill an RL environments engineer role at a vertical startup like Halluminate?

Refolk's index puts the frontier-caliber pool at 693: US professionals holding a Research Engineer, Research Scientist, or Member of Technical Staff title who also list Reinforcement Learning as a skill. The broader "has RL on the resume" pool is 18,725, but only 3.7% clear the frontier title bar. For a nine-person company hiring against four frontier labs, that 693 is the honest universe.

Where should an AI evals recruiter focus first on a Halluminate-style mandate?

Start with the acquired-labeling-vendor cohorts, not the frontier labs. Sepal AI (acquired by Mercor February 2026) and Deeptune (acquired July 2026) each ship a nameable group of engineers with production RL environment experience and recent liquidity. Add Capital One Labs alumni from the last 24 months and Prime Intellect Environments Hub contributors with merged work.

What makes the finance LLM engineer sourcing problem different from a generic ML sourcing problem?

The scarce atom is dual pedigree, not single skill. Halluminate's Westworld benchmark required 88 PE transaction tasks reviewed by practicing deal professionals, which means every useful hire either shipped at a frontier lab and understands DCFs, or spent years in M&A and can also code. Standard Boolean strings return bankers who took a Coursera class and ML engineers who summered at a bank in 2014. Neither is the hire.

Will Halluminate's customers try to block its hiring from their teams?

Almost certainly. In a 25-profile sample of the 693-person pool, 13 people sit at OpenAI, Anthropic, Google DeepMind, or Meta, which overlaps directly with Halluminate's four frontier-lab customers. Expect non-poach language and warrant provisions in vendor contracts to surface as a story inside twelve months, which gives sourcers a narrow window to run before the paperwork catches up.