loading...
Halluminate builds RL training environments for AI agents doing private equity diligence – and is already profitable with nine employees.
ENTRY ANGLES
build RL training environments for legal knowledge work · create expert network for domain-specific AI benchmarks · license domain training data to AI labs · build vertical evaluation infrastructure for consulting/accounting
VERTICALS
CAPABILITIES
domain expert recruitment and compensation, data quality verification, RL environment engineering, benchmark design, AI lab partnerships
The most capable AI systems in the world are being evaluated with benchmarks that were never designed for serious work. Multiple-choice math tests and coding puzzles tell you something about raw capability; they tell you almost nothing about whether a model can complete a 300-step private equity diligence process using real financial data and undocumented institutional conventions.
Halluminate builds what comes next. The company creates RL training environments and benchmarks for AI agents doing knowledge work in financial services: investment banking, private equity diligence, M&A. Its flagship product, the Westworld Finance Diligence Bench, contains 88 problems drawn from anonymized real private transactions, each requiring agents to navigate dynamic desktop environments over hundreds of steps. When Halluminate tested seven leading models against the bench in August 2026, the best performer scored 51% – more capable than a first-year analyst, well short of a senior associate.
Founded in 2024 by Jerry Wu and Wyatt Marshall, both Cornell computer science alumni, the company has raised $38.5M total, with a $30M Series A led by Oak HC/FT in October 2026. It claims profitability at a mid-eight-figure annualized revenue run rate with approximately nine employees. Four of the five leading closed-source US AI labs are paying customers.
The hidden business here isn't the benchmark. It's the training data. Getting an AI agent to perform consistently on a financial diligence task requires not just benchmark scores but ground-truth demonstrations of how expert humans actually approach the work. Halluminate's expert network collects those demonstrations. The result is a dataset that can only be created by finance professionals and is directly usable as reinforcement learning training signal.
Scale AI built a $14B business on the insight that data labeling – structured, domain-specific data creation – is a real infrastructure layer in AI development. Scale's moat wasn't its technology; it was the trained workforce and quality systems that made its data trustworthy at scale. Halluminate's moat is narrower and deeper: finance domain expertise encoded into environments that model developers can't cheaply replicate internally.
The revenue-per-employee figure that results – mid-eight-figures divided by roughly nine people – is extraordinarily high for any software company and suggests the training environment business is genuinely scarce infrastructure at this moment in the AI development cycle.
The adjacent opportunity is direct and well-defined: legal, healthcare, accounting, engineering. Any knowledge-work domain with high complexity, active AI deployment pressure, and institutional data that is expensive to access or difficult to structure is a candidate for the same infrastructure. The challenge is that each domain requires rebuilding the expert network and methodology from scratch, which is what limits expansion speed more than capital.
The specific entry angle for a builder: the Halluminate model works because finance professionals can be compensated to provide training demonstrations, there is a measurable standard of correctness (whether the diligence conclusion was defensible), and the buyer – AI labs – has both the budget and the technical sophistication to use the data effectively. Find a knowledge work domain that meets all three conditions. Legal due diligence and accounting audit processes are the closest matches; medical diagnosis is the most obvious adjacent category but comes with more regulatory complexity.
What Halluminate has proven is that the business model works before the competitive moat is established. The race to replicate it in adjacent domains will produce several valuable companies. The narrow window before cloud providers build generic alternatives is the constraint to design around.