A local investigation console for the Hacker House Goa benchmark. Casework combines TigerGraph graph traversal, native vector search, official TigerGraph MCP, a local language model, and deterministic approval rules.
- Investigates a supplied trigger using customer history, explicitly attributed card transactions, device neighbors and historical cases.
- Detects testing sequences, unusual online activity, new-device signals, out-of-region activity, mixed-channel takeover hypotheses and coordinated new-device/proxy motifs.
- Records uncertainty and explicit simulated verification responses; recommendations never execute banking operations.
- Produces the required case JSON, initial/final action routes and policy-required SAR draft.
- Saves the case and versioned investigation memory to TigerGraph. Future investigations retrieve earlier generated memory with its simulation disclaimer.
- Shows evidence, graph relationships, actions, SAR drafts and execution traces in a browser.
Probabilities are transparent heuristics, not calibrated estimates. Pattern hypotheses are not confirmed outcomes. No hidden benchmark labels or public Kaggle outcomes are used.
Requires Python 3.11+, TigerGraph Savanna with the supplied dataset loaded, and Ollama with llama3.2:3b and nomic-embed-text.
python -m pip install -r requirements.txt
Copy-Item .env.example .env
# Fill in your own TigerGraph host, graph and secret.
python -m scripts.deploy_agent
python -m scripts.deploy_memory
python -m scripts.agent.generate_cases
python -m uvicorn ui.server:app --host 127.0.0.1 --port 8000Open http://127.0.0.1:8000. On an existing prepared graph, skip deployment and ingestion.
For a fresh graph, follow data loading and deploy the schema first. Put the organizer-provided files in data/raw/; the large dataset and credentials are excluded from Git.
python -m scripts.agent.generate_cases --case HHG-010 --response denied
python -m scripts.agent.generate_cases --case HHG-010 --response confirmed
python -m scripts.agent.generate_cases --case HHG-010 --response no_reply_24hThese are assumptions for the challenge, not received customer replies. Rerunning a case updates its current answer and appends a versioned memory record. Run the default benchmark again to restore pending-response answers.
flowchart LR
UI[Local console] --> A[Bounded investigator]
A --> MCP[Official TigerGraph MCP]
MCP --> G[As-of GSQL context]
MCP --> V[Native vector search]
G --> F[Features and graph motifs]
V --> L[Grounded LLM planner/reviewer]
L --> P[Policy and approval routing]
F --> P
P --> J[Validated case and SAR JSON]
J --> M[Graph case and versioned memory]
M --> V
The model chooses a retrieval query and an evidence request and reviews retrieved evidence. It cannot execute arbitrary GSQL or financial actions. Policy code controls action routing. Original historical labels remain separate from generated memory.
python -m pytest -q -p no:cacheprovider
python -m scripts.audit_submissioncases/HHG-001.jsonthroughcases/HHG-020.json(20 required answer files, at repository root)- Model configuration
Execution results and publication status should be checked against the latest audit. Unit tests establish specific behavior, not hidden-label accuracy.
The dataset does not give a card ID for every transaction. Only explicit historical and trigger anchors form card-transaction edges. Customer history is not silently relabeled as card history. Generic device profiles and email domains are not unique people or merchants. Merchant identity and settlement status are absent, so recurring-merchant and cleared-purchase rules require additional verified evidence.
Savanna auto-start and auto-suspend should remain enabled. Credentials stay in .env; the console binds to loopback and should not be exposed publicly without authentication.
IEEE-CIS/Vesta dataset with challenge additions by TigerGraph. See the organizer-provided README for the authoritative benchmark and policy. Regulatory context is attributed in the retrieved policy documents.