Salman Farcy Peer Data Scientist
Resume
All work

Sentinel AI

Agentic AI · Financial crime June 2026

Agentic fraud and AML alert review that measures the recall cost of every case it closes.

Project links

GitHub

Results

38% False positives cleared At a measured, published recall cost.
91% Overturn precision When it contradicts the detection model, it is right 91 percent of the time.
88% Step recall Against a per domain answer key of required investigative steps.
633 Held out cases Frozen split, spent once on the final system.

Skills

  • Python
  • LangGraph
  • XGBoost
  • NetworkX
  • OpenAI API
  • Pydantic
  • FastAPI
  • Docker

About this project

Sentinel AI adjudicates the fraud and AML alerts a detection model has already raised, where roughly 90 percent are false positives. It investigates each alert with deterministic tools, writes a cited report, and stops for a human to sign. The language model chooses which tool to call, but never emits a fact, a number, or a verdict. Every figure comes from tested Python, and a mechanical citation checker rejects any sentence not backed by collected evidence.

Built across two domains on one engine: fraud on card payments and AML on a crypto transaction graph. The eval harness reports the recall cost of every case it closes, which is the number vendors in this category do not publish.