loading...
Theorem uses formal verification to mathematically prove AI-written code is correct before it ships — addressing the oversight gap that fast AI coding has created.
ENTRY ANGLES
Own the default CI/CD verification gate for AI-generated code (GitHub Actions, GitLab CI) · Build liability documentation product for regulated-industry compliance buyers (financial services, medical devices) · Partner with AI coding tools to bundle verification into code generation workflows
VERTICALS
CAPABILITIES
Formal methods expertise, AI proof model training, CI/CD integration, Regulatory compliance knowledge
Jason Gross spent his PhD at MIT building verified cryptographic code that now secures HTTPS — the protocol behind every browser padlock across the internet. Rajashree Agrawal trained the AI models that power Theorem's verification pipeline. Together they are trying to solve the problem their careers intersect on: AI can generate software at a rate that makes human review impossible. Formal verification can prove it's correct. Neither has ever been combined at production speed.
The mechanism Theorem is building around is not new. Formal verification — mathematical proof that a piece of software behaves as specified, without relying on testing — has been a known technique since the 1960s. It has been applied to aircraft control systems, cryptographic protocols, and hardware designs. The reason it never achieved mass adoption in commercial software is that writing a formal proof historically required more effort than writing the code being verified. That constraint was practical, not theoretical.
AI coding changes the economics. If software is being generated automatically, the proof of correctness can also be automated — or at least heavily assisted. Theorem combines formal methods with AI models trained specifically to generate and check proofs, reducing what once took years of PhD-level engineering to something the company says can be completed in weeks or days. The $6 million seed round was led by Khosla Ventures, with participation from Y Combinator, e14, SAIF, and Halcyon. The company emerged from YC's Spring 2025 batch.
The problem Theorem is addressing is what the founders call an oversight gap: the rate of AI code generation has outpaced the rate of human code review, and the gap is widening. AI coding tools can generate production-quality software in hours that would have taken senior engineers days. The reviewers — humans whose review speed has not changed — are falling behind. For most software, a bug is a bug. For financial systems, power grids, and medical devices, it is a liability event.
The regulated industries are the obvious starting point because they already have mandatory validation requirements. A software change to a trading system requires documented evidence that it behaves as specified. A firmware update to a medical device requires FDA review of the validation methodology. These processes were built around the assumption that software changes were infrequent and that qualified humans could review them all. AI coding breaks that assumption. Theorem's value proposition in those industries is not that its output is better than human review — it's that its output is auditable, mathematically grounded, and available at the speed at which AI-generated code actually ships.
The HTTPS connection to Jason Gross's prior work is not incidental. The cryptographic protocols that Gross helped formally verify are among the most consequential pieces of software in existence, and they've been running correctly in production for years without bugs in the formally verified portions. That's the existence proof that formal verification at high stakes actually works — and that it can be maintained over time.
The most direct integration point is inside the CI/CD pipeline. Currently Theorem is a standalone verification tool. The natural default position is a GitHub Action or GitLab CI check that runs automatically whenever an AI-generated PR is submitted for review: it either certifies that the specified properties hold, or it flags the exact properties that couldn't be proven, with a structured report that goes into the PR review. That turns formal verification from a deliberate choice into a default gate — the same path that security scanning and linting took to becoming standard practice. The company that owns the default CI integration for AI-generated code verification owns the first position a developer encounters.
The second market is liability documentation for regulated industries. A financial services firm or medical device manufacturer doesn't need to understand the mathematics of formal verification to value a report that states: "These ten behavioral properties were mathematically proven for this code, at this commit hash, using these formal axioms, by Theorem on this date." That document is a compliance artifact — proof of a validation process that satisfies regulatory requirements for documented software testing. Selling that artifact to compliance and legal buyers, at a price point that reflects regulatory value rather than developer tooling value, reaches a different buyer with a different willingness to pay than the CI integration does.