Analysis
Proximal, a year-old New York City startup, raised a $15 million seed round led by General Catalyst at a $300 million valuation, with SV Angel, Go Global Ventures and Chemistry also participating, The SaaS News reported. The round's individual backers are notable in their own right: Liam Fedus and Kevin Weil, both OpenAI executives, and Erik Bernhardsson, founder of Modal Labs.
The company describes its product as automating the post-training feedback loop across complex operational domains. In practice, that means processing unstructured real-world data and AI agent execution traces to build evaluation environments that surface model reasoning failures and generate targeted datasets for performance tuning -- the same category of work Preference Model and Scale AI operate in, but aimed specifically at post-deployment feedback rather than pre-training alignment. Proximal plans to use the funding to expand its engineering and research teams and to push into computational drug discovery, custom semiconductor design, and legacy software refactoring for critical public infrastructure -- three technical domains with very different data, regulatory and customer dynamics.
“The company describes its product as automating the post-training feedback loop across complex operational domains.”
A $300 million valuation on a $15 million seed is a striking multiple even by this year's AI-infrastructure standards, and it says more about investor appetite for anything touching post-training evaluation than it does about Proximal's current revenue, which the company has not disclosed. Mercor and Braintrust have both raised at escalating valuations in adjacent AI-evaluation categories over the past year, suggesting investors are pricing the sector's addressable market, not any single company's traction, into these marks.
The risk in Proximal's roadmap is breadth: evaluation infrastructure that works for frontier-model reasoning failures doesn't automatically transfer to drug discovery or chip design, where domain expertise and regulatory context matter as much as the underlying ML tooling. A platform trying to generalize across all three at once, this early, is making a bet that its core infrastructure is domain-agnostic -- a bet that's unproven until Proximal names a customer outside pure AI-model evaluation. What to watch: whether the Series A, whenever it lands, cites revenue from even one of those three new verticals.