🔎
Select a case
Choose an alert on the left to open its investigation.
TigerGraph × Hacker House Goa 2026
Fraud investigations that know when they don't know enough.
An AI agent that works a card-fraud alert like an analyst would. It searches the transaction graph, recalls similar past cases, asks the customer a question only when it's worth it, and explains every decision it makes.
- Graph + vector search in one database
- Agent reaches the graph through TigerGraph MCP
- Humans approve every block and filing
- Every decision is audited
–Alerts investigated
–Confirmed as fraud
–Changed decision after evidence
–Regulatory reports (SARs)
–Transactions in the graph
How it works
From alert to explained decision in five steps
The AI plans and writes the explanations. The decisions come from the graph and the bank's policy, in code the AI can't touch.
- ⚡
Alert arrives
From the bank's risk model, a customer saying "that wasn't me", or an analyst's question.
- 🕸️
Search the graph
Cards, devices, regions and past behaviour, all capped at the alert's timestamp. No peeking into the future.
- 🧠
Recall past cases
Finds similar closed investigations by meaning and by shared cards or devices, in one TigerGraph query.
- ❓
Ask only if worth it
Unsure? It recommends a safe first step, then picks the question with the most information per unit of customer hassle.
- ✅
Decide & explain
Final actions with the policy rule cited. Blocks and filings wait for a human to approve.
Start here
Three cases that show what it can do
Why TigerGraph
Some fraud only exists as a shape in the graph
🔗
Hybrid memory in one query
Vector search over 5,565 past case narratives, fused with a traversal to the cases that touched the same cards and devices, in a single GSQL query.
🧮
Graph algorithms find the ring
WCC and PageRank from TigerGraph's GDS library run on the anonymous-proxy slice of the card network. The HHG-014 ring is its own 54-card component; ordinary cards stand alone.
🧩
Every case becomes part of the graph
Each investigation is written back as a FraudCase linked to its evidence, decisions and actions, so the next analyst can find it.
🛡️
Guardrails the AI can't argue past
Probabilities, actions and approval routes come from deterministic code. A test fails the build if that code ever imports an AI model.
See the reasoning behind every call
Open any case to see the evidence, the graph, and exactly why the agent stopped.
Case inbox
20 alerts, each fully investigated. Pick one to see how the agent reached its decision.
Scoreboard
All 20 decisions side by side, each checked against the raw data and the bank's policy.
Audit = eval/audit_answers.py: data integrity, approval routes (§2), case and SAR rules (§3a), exposure (§4), R1, R2, R3, R7, stopping (§6) and graph read-back. There is no answer key, so this proves consistency, not correctness.
Proactive monitoring
Alerts nobody raised. The agent swept November–December on its own and investigated what it found, end to end, under the same policy.
Approval queue
The agent carries out low-risk actions itself. Blocking a card or filing a regulatory report waits here for a person.
Insights
How the agent was calibrated, and the bank policy it has to follow.