Company profile · 4 min read
Cajal: AI-assisted formal verification for scientific systems
Cajal uses Lean-based formal verification and AI mathematicians to discover and validate tools for quantum computing, finance and scientific research.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Cajal is trying to make formal verification useful outside the small community that can write Lean proofs by hand. Its starting buyer is a research or engineering team in a domain where a plausible answer is not enough: the result must be machine-checked against a formal statement.
What it does
The YC profile describes AI mathematicians that discover and formalize tools in Lean, beginning with quantum computing and finance. Its system, Tau, is described as a multi-agent system that takes a research direction, formalizes applied mathematics and returns results checked by Lean’s type-checking kernel. Cajal also says it partners with frontier AI labs and research institutes on datasets, evals and reinforcement-learning environments.
The current Cajal site positions formal verification as infrastructure for mathematics and software. It highlights Talos, an open-source Lean-WASM interpreter, and a Tau-prover project. The site’s claim that Tau came third on a difficult benchmark is company-published; a rank is useful context, not proof that a customer’s system is verified.
Why I’d look closer
The product’s advantage is a hard boundary around correctness. Lean can check whether a formal proof type-checks, which is much stronger than asking a model to sound convincing. Founder context supports the technical focus: YC describes Pedro Nobre as working on formal verification and AI, and Luke Johnston as a machine-learning and neuroscience researcher from Oxford, Cambridge and UCL environments.
The tradeoff is formalization cost. A theorem or software property that has not been expressed in Lean cannot be checked by Lean, and a proof of the wrong abstraction can still miss a real operational requirement. Teams also need to understand what Cajal is delivering: a proof, a reusable library, an algorithm, a dataset or research partnership.
What I’d ask
What claims can Tau formalize in our domain today? Who defines the specification and translates our informal requirement into Lean? Can the proof, dependencies and generated tool be reproduced and maintained as the code changes? For a quantum or financial system, what remains outside the formal model?
My editorial take
Shortlist Cajal when correctness is an explicit product requirement and your team can invest in the specification layer. Start with one bounded property and require a human domain expert to review the formalization, not just the green check. The company’s opportunity is real infrastructure; the adoption question is whether the proof boundary matches the risk you actually need to control.
Quick facts
| Field | Sourced detail |
|---|---|
| Product | AI-assisted formalization and verification in Lean, including Tau and Talos |
| Buyer | Quantum, finance, scientific-computing and frontier AI research teams |
| Public asset | Talos open-source Lean-WASM interpreter |
| Pricing | Not published in the checked pages |
| Main question | Does the formal specification cover the failure that matters in production? |
Sources checked
| Source | Checked |
|---|---|
| YC company profile | 2026-09-19 |
| Cajal homepage | 2026-09-19 |
| Talos on GitHub | 2026-09-19 |
| Cajal YC launch | 2026-09-19 |
Cohort context
Cajal is listed in Winter 2026. In our 2026-09-18 directory snapshot, 126 of 199 listed companies in that cohort have YC’s primary industry label B2B (63.3%). This is a current-directory comparison, not an original intake count or a performance ranking. Nine-cohort dataset.
Public website snapshot
Observed 2026-09-19T16:20:04.824Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.
| Signal | Homepage observation |
|---|---|
| Product description metadata | Observed |
| Canonical link | Observed |
| H1 or H2 heading | Observed |
| Typed structured data | Observed |
| Docs/developer link | Not observed in this response |
| Pricing link | Not observed in this response |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |
Public observations · Collection method. Missing links here do not establish that a capability or file is absent elsewhere.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗