FraudGraph

FraudGraph dashboard showing the Jac Agent Walkers panel — all four walkers for account ACC-4CE8E824.
Account Risk Distribution scatter plot showing fraud scores across 17 accounts, color-coded by risk level (low → critical).
Transaction Feed showing high-confidence flagged transactions with reasons like geo anomaly, velocity spike, and device mismatch

💡 Inspiration

Fraud detection today is mostly reactive — rules fire after the damage is done. We wanted to build something that thinks like a fraud investigator: one that follows the money, spots patterns across accounts, and makes judgment calls autonomously. When we discovered Jac's Object Spatial Programming model, it clicked — walkers traversing a graph are investigators walking a case file.

🔨 How We Built It

FraudGraph is built around four Jac walkers that traverse a live transaction graph, each with a specialized role:

Walker	Role
`FraudDetectionWalker`	ML-scores every transaction using Isolation Forest
`AnomalyPatternWalker`	Detects velocity spikes, geo anomalies, merchant risk
`EscalationWalker`	Uses `by llm()` to recommend block / freeze / verify
`SynthesisWalker`	Uses `by llm()` to write an executive fraud report

The backend is Python FastAPI with SQLAlchemy + SQLite, exposing walker endpoints that the React + Tailwind frontend calls in real time. The walkers genuinely traverse account → transaction nodes, collecting evidence before making decisions — this is not a UI wrapper around a single API call.

🚧 Challenges We Faced

Jaclang versioning — the by llm() syntax and walker APIs evolved rapidly; getting a stable version compatible with Python 3.10 on Windows took significant debugging.
Graph ↔ REST bridge — wiring Jac walker state back into FastAPI responses cleanly required building a custom jac_runner.py bridge layer.
Realistic fraud simulation — generating seed data that actually exercises all four walkers meaningfully (velocity attacks, cross-border anomalies, merchant clusters) took careful design.

📚 What We Learned

Object Spatial Programming is a genuinely different way to think about agentic AI — agents live in the graph rather than being called externally.
by llm() makes LLM-powered decisions feel like first-class language features, not bolted-on API calls.
Building a graceful simulation fallback was the right call — it kept the demo reliable regardless of environment issues.