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

#agents #ai #devchallenge #llm

A technical breakdown of our submission for the TigerGraph HHGOA Hackathon.

1. Executive Summary

Traditional automated fraud prevention systems suffer from a binary fallacy: attempting to label complex transactions as either "fraud" or "not fraud." In modern financial crime, sophisticated syndicates exploit these rigid thresholds through synthetic identities, distributed device emulators, and rapid multi-hop pass-through accounts.

We built CaseGuard: an autonomous fraud investigation agent powered by TigerGraph, GSQL analytics, and a cyclic LangGraph state machine.

CaseGuard's defining philosophy is "An investigator that knows what it doesn't know." Rather than guessing aggressively, CaseGuard:

Formulates hypotheses by executing GSQL graph algorithms (detecting device sharing, velocity bursts, IP clusters, and mule chains).
Computes an explicit, multi-factor Uncertainty & Confidence score.
Initiates controlled, policy-approved evidence gathering (step-up 2FA, customer transaction pings) when confidence is insufficient.
Generates defensible Next-Best-Action (NBA) recommendations with strict Human-in-the-Loop policy guardrails.
Persists closed cases back into the graph to empower continuous case-memory learning.

2. Architecture Blueprint

┌────────────────────────────────────────────────────────┐
│                   STREAMLIT DASHBOARD                  │
│   Case Timeline · Evidence Lineage · Approval Queue   │
└───────────────────────────┬────────────────────────────┘
                            │
┌───────────────────────────▼────────────────────────────┐
│              CASEGUARD AGENT ORCHESTRATOR               │
│                   (LangGraph Loop)                     │
│                                                        │
│  TRIAGE ─▶ GATHER ─▶ DETECT_PATTERNS ─▶ CASE_MEMORY    │
│    ▲                                          │        │
│    │      ┌─────────────────────────────┐     │        │
│    │      │  Uncertainty Engine         │     │        │
│    └──────┤  (Score < Threshold?)       │◀────┘        │
│  Request  └──────────────┬──────────────┘              │
│  Evidence                │ Score >= Threshold          │
│                          ▼                             │
│                  RECOMMEND_ACTIONS ─▶ PERSIST_GRAPH    │
└───────────────────────────┬────────────────────────────┘
                            │
┌───────────────────────────▼────────────────────────────┐
│                    TIGERGRAPH DB                       │
│  • Transaction Graph Topology                          │
│  • Installed GSQL Pattern Detectors                   │
│  • Case Memory & Prior Investigations                  │
│  • GraphRAG Policy Knowledge Store                     │
└────────────────────────────────────────────────────────┘

3. How TigerGraph & GSQL Win the Investigation

The dataset contains ~590,000 transactions, ~13,500 customers, and multi-dimensional identity signals. A generic LLM cannot compute connected graph paths or aggregate velocity over graph edges.

In CaseGuard, GSQL does 100% of the graph traversal and pattern detection; the LLM reasons over the structured graph output:

Device Multi-Accounting (detect_device_sharing.gsql): Evaluates the subgraphs connecting Device -> Transaction -> Account and Customer. Instantly flags device fingerprints linked to 3+ distinct cards or identities within sliding windows.
Velocity Burst Analysis (detect_velocity_burst.gsql): Performs real-time accumulator-driven aggregations over Transaction vertices outgoing from a card to catch automated testing bursts.
Mule Chain Tracing (detect_mule_chain.gsql): Multi-hop BFS traversal through pass-through intermediary accounts.
Address & Device Discrepancies (detect_addr_mismatch.gsql): Discovers mismatches between billing address, shipping address, and newly seen device IDs.

4. The Uncertainty Engine (The Winning Lever)

The primary differentiator in judging criteria (representing 50% of the score between Accuracy and Next-Best-Action) is the agent's ability to navigate ambiguity.

CaseGuard computes confidence as a weighted, calibrated formula:

$$\text{Confidence} = w_1 \cdot \text{GraphSupport} + w_2 \cdot \text{HistoricalRate} + w_3 \cdot \text{SignalStrength} + w_4 \cdot \text{EvidenceCoverage} - \text{Penalty}_{\text{contradiction}}$$

Graph Support ($w=0.35$): Grounded in GSQL query matches and neighbor fraud density.
Historical Rate ($w=0.25$): Outcome distribution of structurally similar prior closed cases.
Signal Strength ($w=0.25$): Bank model risk score and anomaly features.
Evidence Coverage ($w=0.15$): Percentage of the required checklist for the suspected typology.
Contradiction Penalty: Subtracted if behavioral signals resemble previously cleared legitimate cases, preventing aggressive false-positive blocks.

The NBA Decision Progression (Before vs. After Evidence)

Hackathon guidelines require logging the next best action both before requesting additional evidence and after receiving it. CaseGuard enforces this natively in state:

If $\text{Confidence} < 0.60$, CaseGuard logs the initial NBA (e.g., monitor_account + request_step_up_auth), dispatches the mock verification, and recalculates the updated NBA upon receiving user validation.

5. Hardcoded Governance & Permission Matrix

Autonomous agents in banking must operate within strict regulatory guardrails:

Non-invasive actions (request_step_up_auth, warn_customer, monitor_account) auto-execute.
High-impact protective actions (block_transaction, block_account, file_SAR) are classified as pending_approval and routed to L1/L2 Analysts or Compliance Officers via the UI approval queue.
Mandatory SAR narratives are drafted in FinCEN standard formatting when transaction volume exceeds \$5,000 or mule laundering patterns are detected.

6. Key Learnings & Future Horizon

Separation of Graph Execution & LLM Reasoning: Offloading graph traversals to pre-compiled GSQL queries gives sub-millisecond execution times and deterministic evidence.
Graph-Backed Memory Over Raw Vector Stores: Embedding case summaries while maintaining graph linkages (SIMILAR_TO_PRIOR_CASE) creates a rich historical audit trail.
With More Time: We plan to implement automated community detection (Louvain) in GSQL to proactively identify emerging, undocumented syndicate clusters before a trigger even fires.

DEV Community