Necromancer
I ahve built a agentic fraud detection sysgtem which uses 3 agents to flag the chained fraud transaction which is the ost effective fraud trick.
Core trick
XGBoost flags suspicious account clusters, a Gemini agent runs graph-traversal code to trace fund hops, and a judge rejects weak traces.
Description
I ahve built a agentic fraud detection sysgtem which uses 3 agents to flag the chained fraud transaction which is the ost effective fraud trick.
UPI Fraud Investigation Agents — a multi-agent system that investigates suspected UPI mule-account fraud, and is designed to distrust its own conclusions.
Three agents run in a loop:
Detector — an XGBoost model (trained on the public IBM AML dataset) flags suspicious account clusters. It's tuned recall-first: it catches ~75% of mule transactions, at the cost of about 1 in 4 flags being a false alarm.
Investigator — a Gemini managed agent that writes and runs its own graph-traversal code in a cloud sandbox, tracing the fund flow hop by hop.
Judge — a Gemini agent that critiques the trace against explicit evidence criteria. Weak traces get rejected and sent back for re-investigation (capped at 3 passes); genuinely ambiguous cases get handed to a human rather than guessed at.
README excerpt
# UPI Fraud Investigation Agents A multi-agent system that autonomously investigates suspected UPI mule-account fraud — and refuses to accept its own weak conclusions. - **Detector** flags suspicious account clusters (backed by an AML scoring model — see prior-work note below). Deliberately tuned **recall-first**. - **Investigator** (Gemini managed agent, Interactions API) writes and executes its own graph-traversal code in a cloud sandbox to trace the fund flow hop by hop. - **Judge** (Gemini, structured JSON verdict) critiques the trace against explicit evidence criteria. Weak traces are **rejected and sent back** for re-investigation; genuinely ambiguous cases are **handed to a human** instead of guessed at. The loop is hard-capped at 3 passes. ## Prior work vs. hackathon work **Prior work (declared):** the fraud-scoring model behind the Detector — an XGBoost detector trained on the public IBM AML dataset (HI-Small) — plus that dataset itself. Prior research this system builds on; it was not trained at the event. Held-out test metrics (positive class = fraud, evaluated at natural imbalance — fraud is 0.27% of transactions, so **AUC-PR is the honest headline**, not accuracy): | | | |---|---| | AUC-PR | **0.7548** | | AUC-ROC | 0.9915 | | precision | 0.7483 | | recall | 0.7272 | | F1 | 0.7376 | Threshold (0.8838) is selected on the **validation** split targeting 75% recall; the test set is then scored **once** at that fixed threshold. Test recall lands at