Halluminate raises $30M to train AI on real financial work

Halluminate builds financial-work simulations that AI labs use to test and train their models. The nine-person San Francisco startup measures how systems handle jobs such as acquisition due diligence, then constructs reinforcement-learning environments around the failures. Halluminate announced a $30 million Series A led by Oak HC/FT on October 1, taking its total funding to $38.5 million.

The customer is not the bank or private-equity firm doing the deal. For now, Halluminate sells to frontier model developers that need realistic practice material for agents expected to work across documents, email and changing instructions. Its expert network and verification methods are part of the product, alongside the simulated software environment.

Its August Westworld benchmark put seven frontier models through 88 tasks based on anonymized private transactions and reviewed by practicing deal professionals. One exercise required an agent to revise a statement of work using a 160-file data room, 21 emails across nine threads and four meeting notes. The best average score among tested models was 51%, according to the company’s benchmark.

Those misses become the raw material for Halluminate’s training environments: a model that follows an outdated instruction or drops a required edit can receive targeted practice on the same type of breakdown. CEO Jerry Wu told Fortune that four of the five leading closed-source American AI labs are paying customers. He said the company is profitable; Halluminate did not disclose an exact annualized revenue figure.

Halluminate is deliberately keeping that customer list narrow. Wu said the company wants to work with frontier labs before pursuing enterprise customers, while it deepens its finance expertise and expands toward adjacent kinds of knowledge work. That leaves enterprise sales for later and ties the product roadmap closely to the models those labs are developing.

The tests also have to get harder as the models improve. Wu estimates that Halluminate must roughly double the complexity of its environments every six to eight months, whether through longer task sequences, harder reasoning or larger sets of files.