The problem Bridgewatch solves

The Gap: $15.8B reached sanctioned entities last year. Current tools catch known bad actors — but when a fresh wallet bridges funds from China's digital yuan system, Chainalysis sees nothing unusual. The origin is invisible.

Our Solution: First compliance tool that detects behavior, not just addresses. We apply NLP to transaction calldata — like detecting an accent even when you've never heard the speaker before.
How it works:

Layer 1: On-chain OFAC screening (Merkle proofs, trustless)
Layer 2: ML flags suspicious patterns — institutional timing, gas behavior, origin obfuscation

Why now: GENIUS Act (July 2025) mandates stablecoin compliance. $312B market needs this.

Challenges I ran into

1. False Positive Calibration
The bug: First live test against Ethereum mainnet — 19 out of 20 transactions flagged as suspicious.
What happened: Unknown function selectors were scoring +15 points automatically. The Across Protocol relayer function executeRelayerRefundLeaf wasn't in our whitelist, so every legitimate relayer transaction got flagged.
The fix: Realized this is actually correct behavior for a compliance system — flag unknown patterns, build whitelist over time. Added Across selectors to known list. Understood the difference between "false positive" and "conservative by design."

2. The Cold Start Problem
The challenge: ML needs training data. But there's no public dataset of "CBDC-origin transactions" — that's literally the gap we're filling.
The solution: Generated synthetic transactions mimicking mBridge patterns based on public documentation. Combined with real bridge traffic labeled as "normal." Not perfect, but enough to prove the concept works.

Use of AI tools and agents

Core ML Pipeline:

NLP Feature Extraction — Analyzes transaction calldata patterns using text analysis techniques adapted for blockchain data
Explainability Layer — Shows which features drove each risk decision

Development:
Cursor and Claude assisted with code implementation and debugging throughout the hackathon.