I build agent products that stop before they do damage.
Product manager in supply-chain finance. My work is mostly about the gap between what a model can draft and what a regulated business can act on — policy packs an agent must cite, approval gates it cannot route around, and evidence honest enough to argue with.
Work
KYC Checker Agent
Compliance review · live demoAn agent that reads a KYC case against fifty rules, proposes the checks it needs, and stops. A human approves, edits or declines; only then does the server run anything. Graded on 6 cases and 47 written-in-advance expectations.
Stablecoin receivables platform
Trade finance · private repoThe product the agent lives inside: suppliers submit invoices, funders finance them in USDC, and every status change is re-derived from the chain before the database is trusted. Seven roles, and a gate nobody can self-serve past.
Pricing analyst
NextThe same method in a different domain — a rules engine that sets the band, an agent that reasons inside it and never computes the price itself.
How I work
- Write the decision rule before the run. On the model comparison in the KYC case study, the rule I fixed in advance disqualified the model I ended up choosing. That is recorded rather than reinterpreted.
- Counts, not rates. Six hand-graded transcripts do not make a percentage, and quoting one invites comparison with benchmarks measuring something else.
- Publish the limits at the same size as the results. What the evidence does not cover is part of the evidence.