LIVE · PyTorch GNN + FastAPI

Real-time fraud detection
through the graph itself.

Hybrid scoring engine fusing hand-tuned graph heuristics with a PyTorch edge-level GNN. Every transaction is enriched with 14 features derived from live network context — velocity spikes, reverse edges, mule flow, cross-border paths — and scored in milliseconds.

Launch Simulation Score Transaction
Risk Pressure · last hour LIVE
0 High-Risk /h
p95 Score —
Events 0
Alerts 0
Graph Nodes 0 Unique accounts
Graph Edges 0 Observed relationships
Events 0 Processed txs
Alerts 0 Above threshold
High-Risk 1h 0 Score ≥ 0.85
Service CHECKING v0.1.0
01 / Threat Simulator
SYNTH STREAM
Ready · press Run to generate a fraud-laced stream.
// awaiting stream · live alerts will wire a force graph
benign
medium
high
critical
02 / Alert Stream
0 active
min score 0.70
  • // no alerts yet · run a simulation
03 / Single Transaction Scorer
POST /api/v1/score
Live Response —
// submit a transaction to see the explainable risk breakdown
01

Graph Store

NetworkX DiGraph holds accounts and directed money flow. Every event updates degrees, amounts, and first/last timestamps for O(1) lookups during feature enrichment.

NetworkXIn-MemorySliding Window
02

Hybrid Scoring

14 engineered features (velocity, z-score, reverse-edge, mule ratio) scored by a calibrated heuristic AND a PyTorch EdgeMLP. Model uplift protects against false negatives.

PyTorchBCEWithLogitsFeature Attribution
03

Observability

Every request is instrumented: HTTP latency, score histograms, high-risk counters, graph gauges. Prometheus scrapes, Grafana visualizes, alerts flow downstream.

PrometheusGrafanaFastAPI