0dates

Inspiration

Every year, $800 billion to $2 trillion is laundered through the global financial system. Hong Kong - one of the world's top three financial centers - processes trillions in cross-border flows, making it ground zero for this fight. The analysts protecting the system are stuck with spreadsheets, 2D dashboards, and 95%+ false positive rates. Suspicious Activity Reports that should capture criminal behavior take 4-8 hours each to write. The technology hasn't changed in a decade while the criminals have evolved. We built ANGELA to close that gap.

What it does

ANGELA (Anomaly Network Graph for Explainable Laundering Analysis) is an AI-powered 3D graph intelligence platform for anti-money-laundering investigation. It combines:

3D spatial visualization - financial entities exist in 3D space where Y-axis = risk score (suspicious entities physically rise), X-axis = jurisdiction lanes, Z-axis = KYC level. Node size encodes transaction volume, glow encodes risk. This isn't a dashboard - it's a spatial intelligence environment.
Multi-agent agentic investigation - an analyst types one sentence like "find structuring patterns near the reporting threshold." Four AI agents running on AWS Bedrock take over autonomously: the Intake Agent parses intent, the Research Agent queries and ranks entities, the Analysis Agent generates plain-English summaries, and the Reporting Agent compiles an investigator briefing. The analyst typed one sentence. Four agents did the rest.
Automated SAR generation - one click generates a complete Suspicious Activity Report narrative, a document that normally takes compliance analysts 4-8 hours. Our agents produce it in seconds.
Counterfactual explainability - a "What If?" engine autonomously identifies suspicious transactions, removes them, reruns the full risk pipeline, and shows the delta. It answers: "what would the risk be if this entity behaved normally?" That's explainable AI a regulator can trust.
Three independent risk detectors - velocity anomaly scoring against population percentiles, structuring detection for amounts just below the $10,000 BSA reporting threshold, and DFS-based circular flow/layering detection. Fused with explicit, auditable weights.
Natural language query engine - six supported query types, all in plain English. "Show me high-risk entities" → LLM parses intent → deterministic execution → entities highlighted in 3D.
Real-time everything - WebSocket-driven live updates. Inject an anomaly and watch the entity rise, the cluster form, the beacon appear - all streamed in real time.

How we built it

Backend: Python 3.11 + FastAPI for the API and WebSocket layer. The risk detection engine uses NumPy for statistical scoring, depth-first search for cycle detection, and a weighted fusion pipeline. The multi-agent system follows a supervisor pattern with four specialist agents (Intake → Research → Analysis → Reporting), each passing structured output forward. AI calls route through AWS Bedrock via boto3's native Converse API, with LRU caching and retry logic. Data lives in an in-memory DataStore singleton - no external database needed. 3D assets (beacons, cluster blobs) are generated server-side using Trimesh as GLB files.

Frontend: TypeScript + Vite + Three.js for immersive 3D rendering. GPU instancing renders thousands of nodes at 60fps. Composable visualization layers handle nodes, edges, clusters, and loaded 3D assets independently. The UI includes a multi-step upload wizard, entity detail panel with markdown-rendered AI summaries, executive dashboard with KPIs and heatmaps, a time slider for temporal analysis, and an autopilot camera system that flies through investigation targets.

Architecture: Stateless REST + WebSocket communication. Typed end-to-end with Pydantic (backend) and TypeScript (frontend). Docker multi-stage builds for both services. Dual AI provider support (OpenAI-compatible + AWS Bedrock native) switchable via environment variable.

Challenges we ran into

Cycle detection at scale - DFS-based circular flow detection can be expensive. We added depth limits (max 4) and visit caps (max 500 nodes) to keep it fast without sacrificing detection quality.
LLM latency in a multi-agent pipeline - four sequential agent steps with LLM calls add up. We implemented configurable depth profiles (fast/balanced/deep), aggressive caching with composite keys, and proactive cache warmup to keep the experience responsive.
3D graph layout - encoding risk, jurisdiction, KYC level, and volume into a readable 3D space required careful spatial encoding decisions. We iterated on the axis mappings until the layout was both informative and visually clear.
Real-time coordination - keeping the 3D visualization, risk engine, agent system, and UI all in sync required a robust WebSocket event broadcasting system with automatic dead-connection pruning.

Accomplishments that we're proud of

14 distinct AI/ML techniques integrated into a single coherent platform
A genuinely agentic system - not just LLM wrappers, but autonomous multistep investigation with memory, orchestration, and tool calling
Counterfactual explainability - something no commercial AML tool offers today
SAR automation that turns hours of analyst work into seconds
A 3D spatial encoding scheme for financial crime that makes hidden patterns visible

What we learned

Multi-agent orchestration requires careful state management - tracking run IDs, step artifacts, and status lifecycles made the difference between a demo and a system
Explainability is as important as detection - regulators don't trust black boxes, and the counterfactual approach bridges that gap
3D visualization of graph data is surprisingly underexplored in fintech - the spatial encoding immediately reveals patterns that are invisible in 2D tables

What's next for ANGELA

Pilot with a Hong Kong compliance team on real transaction data
Benchmark false positive reduction on labeled AML datasets
Integrate with banking infrastructure - case management APIs, SWIFT message parsing, KYC databases
Build toward a SaaS platform, starting with mid-tier Hong Kong banks and expanding across APAC
Explore patent protection for the 3D spatial encoding scheme and counterfactual risk analysis method