Pramaana’s system generates machine-checkable proofs, flags breached rules, and withholds answers until correctness is demonstrated, enhancing accountability.

The approach relies on formal verification to ensure outputs adhere to codified rules, aiming to move AI from “probably right” to “provably right” in regulated industries.

Founders Ranjan Rajagopalan, Sanjay Ganapathy, and Krishnan Raghavan lead a globally distributed team, with about half the workforce based in India, and a frontier research lab collaborating with IIT Delhi, IIT Madras, UC Berkeley, and Stanford’s Centaur Lab.

The company plans LEAN-style verification frameworks for each industry, supervised by top advisors, to oversee cybersecurity, drug discovery, and tax-law rule formalization.

A key limitation is the specification problem: a correct proof only validates a correctly specified rule set; incomplete or wrong rules yield incorrect conclusions even if the proof is sound.

A San Francisco-based AI startup is ambitiously embedding a deterministic, mathematically grounded verification layer atop conventional large language models to produce machine-checkable proofs and reduce hallucinations in high-stakes domains like law, tax, and drug discovery.

The effort builds on formalization precedents like France’s CATALA and Microsoft Research’s LEAN applications, framing domain knowledge as verifiable code for AI.

Microsoft and academic precedents underpin Pramaana’s approach to formalizing domain knowledge into verifiable code and LEAN-based proofs.

To achieve reliability, the system translates domain knowledge into a formal language and uses a proof engine to verify answers or reveal rule violations before responding.

Pramaana Labs is building sector-specific verification systems guided by expert advisers, including former IRS Commissioner Danny Werfel and academics from IIT Delhi, IIT Madras, and UC Berkeley, to ensure rule bases are accurate and comprehensive.

He argues that the hardest problems are unformalized and become solvable once rules are codified into a formal framework.

CEO Ranjan Rajagopalan emphasizes codifying rules to produce deterministic domain reasoning while preserving clear user communication.