Halluminate Raises $30M Series A for Finance AI Training

Halluminate has raised a $30M Series A to build the training environments and benchmarks that teach AI agents how to perform complex knowledge work. Oak HC/FT led the round, with Y Combinator, Orange Collective, FT Partners, and Heavybit participating. The October 1 financing brings Halluminate's company-reported total capital raised to $38.5M.

The San Francisco startup is beginning in financial services, where a correct answer is rarely enough. An agent may need to reconcile instructions across emails, inspect a data room, update a model, produce a client-ready document, and preserve terms that changed halfway through the assignment. Halluminate turns that moving target into simulated work that a model can practice and that a lab can score.

That matters because professional AI is running into a data problem disguised as an intelligence problem. Public text can teach a model how financial language sounds. It cannot reproduce the private files, changing instructions, tacit judgment, and multi-step consequences that determine whether real work survives contact with a client.

What Halluminate raised and who is involved

Halluminate's $30M Series A was led by Oak HC/FT and included existing investors Y Combinator, Orange Collective, FT Partners, and Heavybit. The company did not disclose its valuation or a detailed allocation of the proceeds. It said the new financing lifts total capital raised to $38.5M.

Co-founders Jerry Wu and Wyatt Marshall started Halluminate in 2024. Wu, the company's CEO, previously led product and research at Capital One Labs, where he worked on an early AI agent for financial services. Marshall, Halluminate's CTO, previously built large-scale data infrastructure and conducted image-model research; the founders met while studying at Cornell, according to Y Combinator's company profile.

The investor thesis is unusually specific. Oak HC/FT argues that the next constraint in knowledge-work AI is the supply of realistic simulations that combine domain expertise, software environments, and reward design. Halluminate is betting that the companies able to rebuild an industry's work inside software will become part of the training stack for frontier models.

Why financial work exposes the gap

Halluminate's Westworld Finance Diligence Bench contains 88 tasks drawn from anonymized private transactions and written or reviewed by practicing deal professionals. The benchmark spans modeling, underwriting, quality of earnings, data preparation, pitch materials, closing, and diligence review. Agents have to operate across office applications, a data room, email, chat, and notes, sometimes over hundreds of steps.

The strongest tested configuration earned roughly half the available credit. That result should not be read as a universal score for finance AI, but it makes the operating gap visible. Models can produce fluent analysis while missing a changed instruction, using superseded information, applying the wrong method, or dropping a requirement before the final deliverable.

Halluminate uses those failures as a map for training. A benchmark identifies where an agent breaks; a reinforcement-learning environment recreates the work so the system can practice, receive a verifiable reward, and improve. The product is therefore less like a static test bank and more like a factory for increasingly difficult professional situations.

The business case behind the round

Fortune reported that Halluminate works with four of the five leading closed-source U.S. AI labs. The company says it reached a mid-eight-figure annualized revenue run rate in 10 months while remaining profitable. Those are company-reported operating metrics rather than audited financial disclosures, but they help explain why a vertical training-infrastructure company could attract a $30M Series A so early.

The concentration also creates pressure. Halluminate is deliberately serving a small set of frontier-model customers before pursuing broad enterprise adoption. Each new model generation raises the difficulty required from the training environments, which means Halluminate has to manufacture more realistic tasks, longer trajectories, stronger verification, and deeper domain coverage fast enough to remain useful.

Finance is a demanding proving ground because the work is both structured and stubbornly human. A model can be graded on whether a spreadsheet ties, a contract preserves the right clause, or a memo follows the latest instructions. Yet constructing those tests requires professionals who understand which details matter and engineers who can turn that judgment into repeatable software.

What the Series A changes

The financing gives Halluminate room to deepen that combination of research, engineering, and domain expertise. The company is hiring across platform engineering, post-training research, finance research, strategy, and operations, and it says it plans to expand from financial services into adjacent knowledge-work verticals. Halluminate has not published a detailed proceeds schedule, so the capital's effect should be measured through the environments, benchmarks, and customer work it produces.

The broader market is moving in the same direction. As generic training data becomes less differentiated, labs are spending more attention on reinforcement-learning environments, verifiers, and expert-generated tasks. The hard commercial question is which parts of that work labs will build internally and which vertical simulations are too specialized, expensive, or operationally awkward to reproduce on their own.

Halluminate has early evidence that finance can support an external specialist. The next stage will demand more than a larger benchmark library. The company has to keep turning private professional judgment into training systems that become harder at roughly the same pace as the agents learning inside them.