CaseGuard: Winning with Uncertainty-Gated Agentic Fraud Investigation on TigerGraph

Sanskreeti Meshram — Thu, 24 Sep 2026 13:58:17 +0000

🚨 Introduction

Traditional fraud systems often make a simple fraud vs. legitimate decision from static thresholds or risk scores.

But real fraud investigations are rarely that simple.

A transaction can look suspicious because of a new device, unusual geography, or rapid small payments — while still being completely legitimate.

That is why we built CaseGuard: an autonomous AI fraud investigator designed to know what it doesn't know.

CaseGuard combines TigerGraph, GSQL, GraphRAG, and agentic reasoning to investigate suspicious transactions, gather additional evidence when confidence is low, recommend the next-best action, generate SAR narratives, and remember previous cases.

🏗️ Architecture

The CaseGuard workflow is:

Alert → TigerGraph Analysis → Fraud Pattern Detection → GraphRAG Policy Grounding → Uncertainty Gate → Evidence Gathering → Next-Best Action → SAR Generation → Case Memory

TigerGraph acts as the cognitive spine of the system.

Our graph contains entities such as:

Customer
Card
Transaction
DeviceProfile
BillingRegion
ClosedCase
InvestigationCase

These entities are connected through relationships such as OWNS, MADE, FROM_DEVICE, BILLED_IN, and INVOLVES.

⚡ How We Used TigerGraph

Instead of asking an LLM to perform complex graph calculations, CaseGuard uses native GSQL queries.

Card Testing Detection

The card_window query analyzes transaction windows to detect rapid micro-authorizations followed by larger spending.

Device Syndicate Detection

The device_neighbors query expands through shared devices to discover connections between multiple accounts.

In benchmark case HHG-014, the graph revealed one Android device connected to 52 customer cards.

Geographic Anomaly Detection

The region_burst query compares transaction geography against a customer's historical behavior to identify suspicious out-of-region activity.

🧠 Uncertainty-Gated AI

One of our main ideas is simple:

If the system isn't confident, it shouldn't guess.

For example, if an alert is based only on a moderate risk score, CaseGuard does not immediately block the customer.

Instead, it follows policy and requests additional evidence, such as:

Customer verification
Step-up authentication
Transaction confirmation

This creates a safer investigation workflow instead of relying on a single signal.

🔄 Next-Best Action

CaseGuard generates recommendations at two stages.

Before Evidence

The system may recommend:

ALLOW + VERIFY_WITH_CUSTOMER

DECLINE + STEP_UP_AUTH

After Evidence

If evidence confirms fraud, the recommendation can evolve into actions such as:

BLOCK_CARD

CREATE_CASE

FILE_REPORT

with appropriate approval routing such as auto, L1, or L2.

🧠 TigerGraph Case Memory

CaseGuard doesn't forget completed investigations.

Every closed case is stored back into TigerGraph using the insert_case_vertex query.

For example, when HHG-012 generated an out-of-region alert, CaseGuard retrieved a previous case containing legitimate travel history for the same cardholder and used that information to clear the alert.

This turns previous investigations into institutional memory.

📄 Automated SAR Generation

When a case meets the required reporting conditions, CaseGuard can generate a structured FinCEN Suspicious Activity Report (SAR) narrative.

This helps investigators move from detection to documentation without manually reconstructing the entire investigation.

📊 Results

We evaluated CaseGuard against all 20 official Hacker House Goa benchmark cases.

Our documented results included:

✅ 100% schema and policy compliance
✅ Correct identification of multiple fraud typologies
✅ Calibrated auto, L1, and L2 approval routing
✅ Case-memory retrieval for previous investigations
💡 What We Learned

The biggest lesson was the importance of separating graph computation from LLM reasoning.

TigerGraph handles the deterministic relationship and graph analysis, while the AI layer focuses on interpreting structured evidence and making policy-grounded decisions.

We also found that storing investigation outcomes directly in the graph provides useful entity-level memory for future investigations.

🚀 Future Improvements

With more development time, we would add:

Real-time streaming using Kafka/Redpanda.
Community detection using graph algorithms such as Louvain or WCC.
Interactive analyst co-pilot with conversational investigation capabilities.
🏁 Conclusion

CaseGuard demonstrates how TigerGraph + GSQL + GraphRAG + Agentic AI can work together to create a more explainable and evidence-driven fraud investigation workflow.

Instead of simply asking:

"Is this transaction fraud?"

CaseGuard asks:

"What evidence do we have, what don't we know, and what should we do next?"

That's the idea behind CaseGuard — an investigator that knows what it doesn't know. 🕵️‍♀️⚡

TigerGraph #GraphRAG #AIAgents #FraudDetection #FraudTech #GSQL #HackerHouseGoa #HHGOA

DEV Community: Sanskreeti Meshram

CaseGuard: Winning with Uncertainty-Gated Agentic Fraud Investigation on TigerGraph

TigerGraph #GraphRAG #AIAgents #FraudDetection #FraudTech #GSQL #HackerHouseGoa #HHGOA