Definite builds AI agents that generate the reports banks and financial institutions file with their regulators. Before any number lands in a filing, our verification engine checks it independently. Every line gets recomputed from ledger records and checked against a versioned rulebook, and wrong outputs go back to the agent with the reason, like compiler errors. They don’t ship.
We met our first year studying computer science at the University of Waterloo, and interned at Meta, Optiver, and BitGo, places where one unverified number in a filing has real consequences. At BitGo, attestations proving customer assets matched the ledger were assembled by hand every period. At Optiver, daily regulatory numbers were rebuilt manually because nothing downstream took them on faith.
Today a controller pulls data from the core system into Excel and maps it to the report line by line. By the regulators’ own estimate, one quarterly filing takes weeks. The software banks buy for this validates formatting, not truth: a loan mapped to the wrong schedule line passes every check it runs, and an executive personally signs the result.
This is hard because verification has to be independent of the model doing the work. Recomputing every number from source records is the only check a regulator or auditor will accept, and it’s why DeepSeek v3 with our engine beats GPT-5 on public reporting benchmarks.
Every regulated institution files these reports, forever. Whoever proves the numbers owns the layer the filings, the audits, and eventually the books themselves depend on. We build it, and the agents that run on it.