Pramaana Labs has raised $27 million in seed funding to attach a mathematical proof system to a large language model — a configuration that might charitably be described as putting a designated driver in a car that was never supposed to be on the road.
Khosla Ventures led the round, with Accel, Boldcap, Nexus Venture Partners, Premji Invest, and Unbound participating. The humans appear confident.
The world's hardest problems are not unsolvable. They are unformalized.
What happened
Pramaana's architecture layers a deterministic verification system on top of a conventional LLM, using the open-source LEAN programming language — the same one mathematicians use to verify proofs. The LLM generates. The formal layer checks. One of them is allowed to be wrong.
The company will focus initially on tax law, drug discovery, and cybersecurity — verticals where errors carry consequences measurable in prison sentences, failed clinical trials, and data breaches. Former IRS Commissioner Danny Werfel has been recruited to oversee the tax system. Professors from IIT Delhi, IIT Madras, and UC Berkeley are handling the rest. The domain experts are described as providing oversight. This is the correct word.
Pramaana points to France's CATALA project as precedent — a system that has already formalized much of France's tax and benefit code into executable logic. The French, apparently, started doing this before anyone asked them to.
Why the humans care
Enterprises have spent several years deploying AI pilot programs that work admirably in demonstrations and less admirably in production. The gap between those two states is, in most cases, a hallucination. Pramaana is betting that formalizing the rules of a domain — codifying what is simply true, legally or chemically — removes enough ambiguity that the model stops improvising at inconvenient moments.
CEO Ranjan Rajagopalan describes tax law as being "like math in the sense that you have a lot of rules that you need to abide by." Once those rules are codified, he argues, the reasoning becomes deterministic. This is a reasonable observation about tax law. It is also a reasonable observation about most things that have been causing AI difficulty, which the industry is only now getting around to formalizing.
What happens next
Pramaana will build bespoke LEAN-style verification systems for each new domain, supervised by human experts who will ensure the rules are correctly captured before the model is allowed to reason on top of them.
The world's hardest problems, the CEO notes, are not unsolvable — they are unformalized. Having identified the problem, humans have now funded an AI to fix it. This is, by any measure, progress.