FraudLens
EXECUTIVE SUMMARY
FraudLens is an agentic fraud investigation system in which a LangGraph agent autonomously investigates flagged card transactions, retrieves graph evidence from a TigerGraph property graph, grounds its reasoning in policy documents, and produces a calibrated assessment that a human analyst then reviews. The system does not merely flag transactions — it assembles a defensible investigation file explaining why each transaction was or was not escalated.
PROBLEM STATEMENT
Card fraud analysts spend most of their time assembling context manually: pulling merchant history, cross-referencing prior alerts, locating internal policy clauses, and reconstructing a timeline before any decision can be made. Existing fraud tooling surfaces alerts but leaves the investigative reasoning to the human, producing inconsistent escalation decisions and audit trails that cannot be reproduced or defended under regulatory review.
INVESTIGATIVE FINDINGS
- ▸Fraud signals are graph-shaped, not row-shaped: shared devices, addresses, and merchants only become visible as relationship paths
- ▸Transaction-level anomaly scores produce false positives without merchant-category and velocity context
- ▸Policy routing is a business rule, not a model output, and must remain deterministic for auditability
- ▸LLM reasoning without retrieval produces confident but unfounded narrative explanations
- ▸Degraded model output must never overwrite a previously valid investigation file
BEHAVIORAL ANALYSIS
The design deliberately separates reasoning from authority. The model is treated as an investigator whose output must be checked, never as a decision-maker, and every escalation path is governed by an auditable rule table. Provider outages degrade the pipeline to heuristics rather than blocking an investigation, and the system is built so a partial result with an explicit quality flag is preferable to a confident result that silently degrades.
TECHNICAL EXECUTION
A Python LangGraph agent runs a twelve-node state machine: an incoming trigger opens the case, TigerGraph is queried over MCP for subgraph evidence, a GraphRAG layer grounds policy clauses using local TF-hash embeddings written to TigerVector, case memory accumulates prior investigations on the same entities, and a calibrated assessment is produced before a deterministic policy engine maps the outcome to one of ten rules (R1–R10). The agent may only recommend an action; the approval route (auto-approve, L1 review, L2 review) is resolved by code, never by the model. The TigerGraph client spawns its MCP server over stdio and parses markdown-fenced JSON, falling back to a direct REST client when the MCP path fails. A NestJS backend and MongoDB store case state, and a Next.js analyst dashboard renders the evidence graph using precomputed d3-force ticks with drag-reheat physics instead of a continuous animation loop. A rate-limit trickle benchmark retries the job queue through provider quota windows while flagging output quality, and a strict output-schema validator enforces the answer-file contract before any file is scored.
SYSTEM ARCHITECTURE
OPERATIONAL FEATURES
- ✓Autonomous multi-step investigation across a 12-node LangGraph state machine
- ✓Graph evidence retrieval over TigerGraph MCP with automatic REST fallback
- ✓Deterministic policy engine routing every outcome through rules R1–R10
- ✓GraphRAG policy grounding without any hosted embedding dependency
- ✓Case memory that recalls prior investigations on the same entities
- ✓Evidence graph visualisation with precomputed ticks and drag-reheat physics
- ✓Rate-limit trickle benchmark with quality flagging to protect valid output
- ✓Strict answer-file schema validator used as a pre-scoring gate
TECHNOLOGICAL ASSETS
CURRENT STATUS
Active development following the TigerGraph HHGOA hackathon build. Twenty scored investigation files (HHG-001 through HHG-020) were produced and evaluated by judges. Current work extends the policy rule library, hardens the MCP fallback path, and widens the graph schema with additional entity types.
CONCLUSION
FraudLens demonstrates that an LLM agent becomes defensible in a regulated domain only when its reasoning is separated from its authority. Grounding investigations in a property graph and resolving every action through deterministic policy code produces output a compliance reviewer can reproduce, which is the actual requirement in card fraud investigation.