Theorem, an AI and programming languages research lab based in San Francisco, raised $6 million in seed funding to build automated formal verification tools that prove AI-generated code is safe before deployment. VentureBeat reported that Khosla Ventures led the round, with participation from Y Combinator, e14, SAIF, Halcyon, and angel investors Blake Borgesson and Arthur Breitman.
Founded in 2025 out of Y Combinator's Spring 2025 batch, the four-person startup aims to close what its team describes as an "oversight gap" created by AI coding assistants. As automated tools generate billions of lines of code annually, traditional human code review cannot keep pace. Co-founder and CEO Jason Gross noted that "If you asked me to review 60,000 lines of code, I wouldn't know how to do it."
Mathematical Proofs for AI-Written Code
Formal verification mathematically proves that software strictly matches a specification. Historically, the practice was restricted to high-stakes domains such as avionics and cryptography because writing proofs required years of specialized labor—often requiring eight lines of mathematical proof for every single line of code.
Theorem applies machine learning models to write and check these formal proofs automatically. The company uses a technique called "fractional proof decomposition" to allocate verification resources proportionally based on component importance rather than exhaustively testing every behavior. The approach identified a bug that bypassed standard testing at Anthropic. In a technical benchmark called SFBench, Theorem used AI to translate 1,276 formal reasoning problems between Rocq and Lean, automatically proving equivalence in a task estimated to take humans 2.7 person-years.
In one client deployment, Theorem took a 1,500-page specification PDF and legacy software prone to crashes and memory leaks, generating 16,000 lines of verified production code. The customer deployed the resulting parser at 1 Gbps without manual line-by-line review. Gross explained, "Now they have a production-grade parser operating at 1 Gbps that they can deploy with the confidence that no information is lost during parsing."
Gross emphasized that as AI hacking lowers attack costs, automated verification provides an "asymmetric defense" for critical software systems. "Now that formal verification is cheap enough, it might be considered gross negligence to not use it for guarantees about critical systems," Gross said.
Leadership and Academic Research Roots
Theorem's leadership brings deep experience in programming language theory and machine learning:
- Jason Gross (CEO) completed his PhD at MIT, developing verified cryptographic code through the Fiat-Crypto project. That code powers HTTPS connections in web browsers including Chrome and Firefox, as well as the Go standard library and Linux kernel. Gross detailed the underlying compiler and partial evaluation optimizations in a presentation titled "Performance Engineering of Proof-Based Software Systems at Scale".
- Rajashree Agrawal (CTO) is a machine learning research engineer focused on training the AI models behind Theorem's verification engine. Agrawal stated, "We're working on formal program reasoning so that everyone can oversee not just the work of an average software-engineer-level AI, but really harness the capabilities of a Linus Torvalds-level AI."
Gross also presented research on measuring model understanding through proof compression at the Topos Institute in "Compact Proofs: Measuring Quality of Understanding with a Compression-Based Metric", outlining methods to prove tight error bounds on machine learning architectures.
Expanding the Research and Engineering Team
With $6 million in fresh capital, Theorem plans to scale its compute infrastructure and expand into robotics, renewable energy, defense, and cryptocurrency. To support this growth, Theorem is recruiting in San Francisco for four key technical roles:
- Senior SWE ($300,000 - $500,000/yr)
- ML Research Scientist ($300,000 - $500,000/yr)
- ML Research Engineer ($150,000 - $250,000/yr)
- Systems Engineer ($150,000 - $250,000/yr)
Across all four open roles, engineers and scientists focus on post-training to make models as capable at program analysis and formal reasoning as they are at writing code. New hires will also construct interfaces for navigating complex implementations of specifications, helping users spot bugs, write design documentation, and execute non-functional requirements via compiler-level reasoning methods.
Open Roles at Theorem
All current roles are full-time and located on-site in San Francisco, CA.
| Role | Location | Work Mode | Salary |
|---|---|---|---|
| Senior SWE | San Francisco, CA, US | Onsite | $300,000 - $500,000 / yr |
| ML Research Scientist | San Francisco, CA, US | Onsite | $300,000 - $500,000 / yr |
| ML Research Engineer | San Francisco, CA, US | Onsite | $150,000 - $250,000 / yr |
| Systems Engineer | San Francisco, CA, US | Onsite | $150,000 - $250,000 / yr |




