Built with Railtracks

yosemite is a fraud detection and compliance intelligence platform for small businesses. It combines rule-based scoring, machine learning anomaly detection, and multi-agent AI to surface invoice fraud, sanctions exposure, and geopolitical risk.

Built for the TD Best AI Hack track: real-time detection of fraudulent transactions, suspicious patterns, and financial anomalies. The AI fraud analysis pipeline is powered by Railtracks — multiple signals are orchestrated into a single structured report with risk level, summary, recommendations, and duration_ms latency. One click → full analysis.

frontend/        Next.js 14 dashboard (TypeScript + Tailwind)
backend/arrt/    Rust (Axum) REST API — fraud scoring, pipeline ingestion, sanctions
ai/              Python (FastAPI) ML sidecar + Railtracks agent pipelines

• AI Fraud Analysis — one-click pipeline fusing multiple signals into a structured FraudReport:
  • Anomaly detection — Isolation Forest on transaction patterns
  • Benford's Law — chi-squared test for manipulated amount distributions
  • Duplicate invoice detection — repeated order IDs and same customer/amount/date charges
  • Graph analysis — transaction graph heuristics (rings, bursty clusters)
  • Behavioral velocity — 24h vs 30d activity spikes per entity
  • Document fraud — Gemini Vision on uploaded invoices/PDFs
  • Railtracks orchestration — fraud_analyst (single-agent) and fraud_coordinator (multi-agent) fuse all signals into a structured FraudReport (risk level, summary, recommendations, duration_ms)
• Sanctions Screening — entity matching against the OpenSanctions dataset
• Geopolitical Risk — country-level risk briefings via LLM

cd ai
./run_local.sh          # creates .venv, installs deps, starts on :8000

cd backend/arrt
cargo run               # starts on :3001

cd frontend
npm install
npm run dev             # starts on :3000

cd ai
source .venv/bin/activate
python check_agents.py

# visualize execution trees
railtracks init
railtracks viz --port 8002

ai/.env

GEMINI_API_KEY=...      # optional, for document/vision analysis

# Required for "Run full AI fraud analysis" — set one of:
OPENAI_API_KEY=...      # OpenAI or any OpenAI-compatible endpoint
HF_API_KEY=...          # HuggingFace inference endpoint

backend/.env

DATABASE_URL=...
HF_API_KEY=...
HF_BASE_URL=...
GEMINI_API_KEY=...
AI_SERVICE_URL=http://localhost:8000