Emerging RL environmental company achieves double-digit profitability in just over 2 years In addition to providing expert data or simulated data to AI/robot companies as previously mentioned, such as 4DLabs @ 4DLabs_Official and Axis Robotics @axisrobotics）， There is also a type of RL environment company (the following chart shows the current market map of this type of company), whether you call it a data company or an RL environment company. Because the core is to provide training data and environment for the head laboratory. These types of companies have good cash flow in the US market. Today I saw another company with less than 10 employees, but it has already achieved profitability. The RL environment he initially worked in was a financial knowledge work environment that was previously difficult to quantify. Halluminate is a 'test room+training ground' for financial knowledge work Open Internet corpus and code questions have made it difficult to push forward the model. Investment banks and private equity firms have long careers, rely on judgment, and are difficult to automatically score: scanning emails, flipping through data rooms, building models, modifying contracts, pitching, and following up on delivery. Their approach is to first adapt a difficult set of papers from real transactions, see where cutting-edge models fail, and then turn these failed patterns into a reinforcement learning environment that can be reset and has verifiable rewards, Sell to the model laboratory for post training. The characteristics are roughly as follows: Vertical and prioritizing finance. Self proclaimed vertical data research lab. The first stop is financial services, followed by consulting, insurance, and operations. The logic is that teaching an agent to do bank/PE due diligence is not the same type of environment as teaching it to write code. Sell papers first, then sell tutoring classes. The Westworld Finance Diligence Bench is a publicly available benchmark consisting of 88 tasks from anonymous real private equity transactions, written and reviewed by professional traders. Intelligent agents can complete mergers and acquisitions due diligence in a dynamic desktop, with trajectories up to several hundred steps. The highest average score of the 7 cutting-edge models is only about 51%. The paper proves the gap, the environment is used to fill the gap, and the customer is mainly the laboratory, not the buyer themselves. Long cycle, verifiable, not a single round of Q&A. In the environment, it is necessary to scan the inbox, assemble data rooms, build models, and produce deliverables. The rewards are a mixture of agentic, discrete, and binary validators. This is closer to real work than 'answering a question correctly', and it is also safer, parallelizable, and reproducible than training agents on real websites. Extremely slim in business. About 9 people, established in 2024, YC S25。 Series A: $30 million, led by Oak HC/FT, with a total investment of $38.5 million. The company claims that four of the top five closed source laboratories in the United States are paying customers, with annual revenue in the middle of eight figures and profitability. Individual co investors from Anthropic, OpenAI, and Meta researchers. official website http://halluminate.ai ， X is @ HalluminateAI