Pramaana Raises $27M to Make AI Provably Right

Pramaana Labs has raised a $27 million seed round led by Khosla Ventures to tackle one of generative AI's most stubborn problems: you can't trust the output. The round also drew Accel, Boldcap, Nexus Venture Partners, Premji Invest and Unbound. The company's pitch isn't a smarter model — it's a way to make a model's answers provably correct in fields where a confident-sounding mistake can cost someone their money, health, or freedom: law, drug discovery, and tax preparation.

The approach pairs a conventional large language model with a deterministic verification layer built on LEAN, the open-source language used to formally verify mathematical proofs. The idea is to encode a domain's rules — tax law, regulatory requirements — into formal code, then check the model's reasoning against it. As CEO and co-founder Ranjan Rajagopalan puts it, "Once you have a codified version of it, the reasoning on top of it starts becoming deterministic." Instead of hoping the model didn't hallucinate, you get an answer that either passes a hard check or doesn't.

1Rules, not vibes

Rajagopalan frames the bet bluntly: "The world's hardest problems are not unsolvable. They are unformalized. Every domain where being wrong can cost someone their health, money, or freedom has rules." The company points to France's CATALA project — which turned the country's tax and benefit system into executable code — as proof the concept works at the scale of a real legal system. To codify those rules, Pramaana is leaning on domain experts: former IRS commissioner Danny Werfel on tax law, and professors from IIT Delhi, IIT Madras and UC Berkeley on cybersecurity and drug discovery. The team is essentially translating expert knowledge into machine-checkable form.

2Why this matters for you

The dirty secret of today's AI is that it's fluent, not reliable — it produces an answer that reads as authoritative whether or not it's correct, and in high-stakes fields that's not a quirk, it's a dealbreaker. Most of the industry is trying to fix this by making models bigger and better at sounding right. Pramaana is doing something different and, I think, smarter: accept that the model will sometimes be wrong, and bolt on a layer that can actually prove whether a given answer holds up against the rules. It's worth being skeptical — this only works in domains whose rules can genuinely be formalized, which is a narrower slice of the world than the marketing implies, and 'verified against codified rules' is only as good as the codification. But the philosophy is the right one. The next leap in useful AI probably won't come from a cleverer model alone; it'll come from systems that can check their own work.