SENTINEL: Real-Time Fraud Detection
Multi-agent fraud detection for Nepal's banking sector. Four specialised agents score every transaction in parallel, a context-aware synthesis layer fuses them, and risky payments go through dual-path OTP that exposes SIM-swap attacks. Built for Track B of the Global IME AI/ML Hackathon 2026.
- Role
- Full-stack and ML engineer: scoring agents, synthesis logic, API, evaluation scripts and the live dashboard.
- P99 scoring latency
- 18.9 ms
- parallel scoring agents
- 4
- automated tests
- 96
- Links
- Live site
01/Overview
SENTINEL scores each transaction through four specialised agents, weights their output by transaction type, and returns one of three verdicts: allow, verify by OTP, or block. The live demo runs entirely in the browser on simulated transactions. It is a research prototype built for a hackathon, not a deployed bank system.
Technical highlight
A FastAPI orchestrator runs the four agents concurrently with asyncio. End-to-end P99 latency is 18.9 ms on the synthetic benchmark, with an AUROC of 0.742 on labelled seeds.
02/The problem
Hackathon Track B asked for fewer false positives, context-aware detection, protection for brand-new accounts, and a defence against SIM-swap fraud.
03/The solution
A graduated verdict band sends mid-risk transactions to OTP instead of blocking them. Per-transaction-type weights drive the synthesis step, six peer cohorts give new accounts a starting model, and independent SMS and email OTP paths flag SIM-swap when one passes and the other fails.
04/The result
Delivered a working end-to-end prototype with a public demo, a benchmark and evaluation suite, and 96 passing tests.
05/Key features
- 01Four agents scored in parallel and fused with transaction-type weights.
- 02Three-way verdict: allow, OTP interlock or block.
- 03Dual-path OTP that turns SIM-swap into a detectable signal.
- 04Live dashboard with simulated transaction stream.
06/Tech stack
Python
FastAPI
React
Tailwind CSS