An LLM architecture for sanctions screening
A two-model pipeline that adjudicates sanctions-screening alerts: it clears the routine false positives and escalates anything that looks like a real hit.
ML engineer & systems builder · San Francisco
I lead Risk Intelligence & Automations at Chipper Cash, where I design and ship the production side of ML: fraud models, LLM pipelines with real evals, and real-time scoring that moves money or stops it. I also write about systems, risk, and the books I cannot stop turning over.
Selected work
A two-model pipeline that adjudicates sanctions-screening alerts: it clears the routine false positives and escalates anything that looks like a real hit.
An LLM-powered pipeline that reads partner fraud reports, resolves identity, and acts in seconds, replacing 30 minutes of analyst transcription with 20 seconds of automation, 24/7.
How we separated genuine super-referrers from referral fraud.
Writing
You have $100K and a coin that lands heads three times out of four. How much do you bet each flip? On the Kelly criterion, ruin, and how to actually get paid for an edge.
Brothers Karamazov
Ivan finds it easy to love mankind in the abstract but struggles the moment that idea takes on flesh and bone. Alyosha shows what it is to love, not within the bounds of logic, but as an act.
Brothers Karamazov
Ivan has said everything is permitted, and yet he cannot live it. He has a conscience he cannot account for. That is the abyss the devil represents.
If you are working on fraud, risk, or an LLM system that has to hold up in production, I would like to hear about it.