# Halluminate

> Halluminate builds RL environments for financial services. We create computer-use and tool-use training gyms that teach frontier AI models to do real financial work — Excel modeling, pitch decks, and the multi-step workflows behind investment banking, private equity, hedge funds, and consulting — plus the benchmarks and evaluations that measure where models succeed and where they break.

Halluminate is an applied AI and data company based in San Francisco. We work directly with frontier AI labs, graduated from Y Combinator (S25), and raised a seed round to scale the team. Our environments run on fully managed infrastructure and are authored with a network of finance professionals — bankers, investors, traders, and operators — who build training problems from real deal work.

Key offerings:
- **RL environments**: Realistic computer-use and tool-use gyms of enterprise and financial software, running on fully managed infrastructure
- **Expert-authored tasks**: Training problems designed by finance professionals from investment banking, private equity, and hedge funds, graded with verifiable rewards
- **Benchmarks & evaluations**: Proprietary and open-source benchmarks that trace agent performance across long-horizon financial workflows

## Research

- [DealTrace](https://halluminate.ai/blog/dealtrace): Staged review of real private-equity deals tracing where frontier models break across a long analytical task; failures concentrate in forecasting, and forecast quality most limits performance
- [Simulated-to-Real Transfer for Web Agents](https://halluminate.ai/blog/sim-to-real-webagents): Does training a web agent on a simulated Best Buy transfer to the real one? A small-scale sim-to-real study using Westworld's Goodbuy clone
- [Westworld](https://halluminate.ai/blog/westworld): Highly realistic simulated websites for benchmarking and training web agents via verifiable rewards
- [BrowserBench](https://halluminate.ai/blog/browserbench): Benchmark revealing how bot detection systems impact browser agent evaluation, with analysis of stealth vs non-stealth performance
- [WebBench](https://halluminate.ai/blog/benchmark): Benchmark for evaluating browser agents on ~2,454 realistic tasks across 452 real-world websites

## Case Studies

- [Yutori](https://yutori.com/blog/introducing-navigator): Post-training the Navigator web agent with Westworld environments
- [Athena Intelligence](https://halluminate.notion.site/Athena-Intelligence-partners-with-Halluminate-to-improve-performance-of-data-extractor-AI-Agent-17d239940a7480628db9d9dd52fe9cda): Partnership to improve AI agent data extraction performance
- [Amplitech AI](https://halluminate.notion.site/Amplitech-AI-partners-with-Halluminate-to-deploy-CI-CD-testing-pipeline-for-AI-Sales-Assistant-17d239940a74806e8d50ca572aae886c): CI/CD evaluation harness deployment for AI sales agents

## Team

- [Jerry Wu, Founder/CEO](https://www.linkedin.com/in/jerry-wu-7814b0100/): PM lead for first generative AI project at Capital One, Cornell CS/ML
- [Wyatt Marshall, Founder/CTO](https://www.linkedin.com/in/wyatt-marshall-a40797198/): Software and data engineer at two venture backed startups, Cornell CS/ML and Philosophy, Milstein Scholar

## Optional

- [Expert network](https://halluminate.ai/experts): Paid program for finance professionals ($100–250/hour) authoring training problems for AI labs
- [Careers](https://halluminate.ai/careers): Open roles — San Francisco, full-time, in-person, visa sponsorship available
- [Contact](mailto:contact@halluminate.ai): Email the Halluminate team
- [Blog](https://halluminate.ai/blogs): Research, announcements, and case studies
- [GitHub](https://github.com/Halluminate): Open source projects and repositories