AI-powered multi-agent platform for banking fraud detection and AI red teaming in FinTech
A Level 3 agentic AI platform (advisory + analysis + execution) that simulates a fraud investigation law firm. Six specialized AI agents — inspired by the characters of Suits — collaborate to detect fraud, ensure regulatory compliance, investigate criminal networks, and red-team AI systems.
Unlike simple LLM wrappers, FraudAI agents execute real tools: analyze transaction datasets, generate compliance reports, build fraud network graphs, and run adversarial attacks against ML models — all inside sandboxed Docker containers.
| Agent | Role | Personality |
|---|---|---|
| Donna Paulsen | Router | Classifies intent, routes to the right specialist |
| Harvey Specter | Transaction Fraud Detection | Direct, confident. Analyzes transactions, detects anomalies, scores risk |
| Louis Litt | AML / KYC / Compliance | Meticulous. Cites exact legal articles, generates SAR reports |
| Jessica Pearson | Fraud Intelligence | Strategic. Network graph analysis, identity resolution |
| Mike Ross | AI Red Teaming | Creative. Adversarial attacks, prompt injection testing |
| Rachel Zane | Data Engineering | Rigorous. ETL pipelines, feature engineering, data quality |
Key components:
- LangGraph StateGraph orchestrates agent routing with conversational continuity
- RAG pipeline with 35K+ Spanish legal documents from the BOE (Official State Gazette)
- Qdrant vector database with hybrid search (dense BGE-M3 + sparse BM25)
- Sandboxed execution via ephemeral Docker containers (cap_drop=ALL, network isolation)
- Groq/Anthropic LLM with multi-provider support and 3-tier routing fallback
- JWT authentication with RBAC per tier (Free/Pro/Enterprise)
- Prometheus + Grafana monitoring with 23 alert rules
| Layer | Technology |
|---|---|
| LLM Orchestration | LangGraph, LangChain |
| LLM Providers | Groq, Anthropic Claude, Ollama (local) |
| Vector DB | Qdrant (hybrid search) |
| Embeddings | BGE-M3 (1024-dim, multilingual) |
| Backend | FastAPI, Python 3.11+ |
| Frontend | Next.js 14, React, Tailwind CSS |
| Auth | JWT (PyJWT) with OAuth2 |
| Sandbox | Docker containers with seccomp |
| Monitoring | Prometheus, Grafana |
| CI/CD | GitHub Actions (lint, typecheck, test, docker, security) |
| Database | SQLite (sessions/feedback), Qdrant (vectors) |
- Docker + Docker Compose
- Python 3.11+
- Node.js 20+ with pnpm
- A Groq API key (get one free)
git clone https://github.com/adrianinfantes/FraudAI-Agent.git
cd FraudAI-Agent
cp .env.example .env
# Edit .env and add your GROQ_API_KEYdocker compose up -d # Qdrant + Ollamauv sync --dev
uv run python -m uvicorn fraudai.api.app:create_app --factory --host 0.0.0.0 --port 8000cd frontend && pnpm install && pnpm dev- Chat UI: http://localhost:3000
- API Docs: http://localhost:8000/docs
- Qdrant Dashboard: http://localhost:6333/dashboard
uv run python -m fraudai.ingestion --mode fullsrc/fraudai/
agents/ # 6 agent definitions, prompts, LangGraph orchestration
api/ # FastAPI endpoints, auth, session management
core/ # Config, database, metrics, tracing
evaluation/ # RAGAS benchmark (50 questions, 5 categories)
ingestion/ # BOE download pipeline, text extraction, chunking
rag/ # Retriever, reranker, prompt templates
tools/ # Sandbox engine, tool implementations
frontend/ # Next.js 14 chat UI with SSE streaming
monitoring/ # Prometheus alerts, Grafana dashboards
docs/ # ADRs, requirements, API reference, runbook
tests/ # 482 unit + 23 live smoke tests
| Document | Description |
|---|---|
| Requirements (F0) | Business, user, system, and ML requirements |
| Architecture Decisions (F0.5) | 6 ADRs: LLM strategy, vector DB, LangGraph, sandboxing, deployment |
| Backlog (F0.5) | 74 user stories, 8 sprints, MVP scope |
| Technical Spec (F2) | Metrics, baselines, SLAs, fairness evaluation |
| API Reference (F6) | All endpoints with curl examples |
| Deployment Guide (F6) | Prerequisites, config, production deploy |
| Operations Runbook (F6) | Common issues, rollback, scaling |
# Unit + integration tests (482 tests, 93% coverage)
uv run python -m pytest tests/ -q
# Live smoke tests (requires running backend)
./scripts/smoke-test.sh
# Coverage report
uv run python -m pytest tests/ --cov=fraudai --cov-report=html- Hybrid LLM routing: Keywords (instant) -> Groq API (200ms) -> Ollama (fallback)
- Conversational continuity: Follow-up messages stay with the current agent
- RAG with auto-filtering: Mentions of specific laws trigger metadata filters
- Security-first sandbox: All code execution in Docker with dropped capabilities
- HITL for red teaming: Mike Ross requires explicit user approval before offensive actions
This is a portfolio project demonstrating AI agent architecture for FinTech fraud detection. It is functional and tested but not production-deployed.
What works:
- All 6 agents respond with domain expertise
- RAG with 35K+ BOE legal documents
- Conversational multi-turn sessions
- Agent routing and override
- JWT authentication
- SQLite persistence
- Prometheus metrics
Adrian Infantes — Senior AI Engineer specializing in FinTech, fraud detection, and AI red teaming.
MIT License. See LICENSE.