Most fraud-agent demos follow a tired pattern: a model spits out a risk score, and an LLM hallucinates a justification for it. A new project on DEV.to by Pranav Ssalian breaks that mold by using TigerGraph and DuckDB to build a fraud investigator that actively argues both sides of a case.

Challenging the Score-First Paradigm

The core insight here is that relying on a single risk score creates a confirmation bias loop. By profiling five provided data files with DuckDB before writing any code, the developer demonstrated that the standard approach often misses nuance. Instead of justifying a pre-determined score, this system uses graph analytics to explore relationships and contradictions in the data.

Graph Intelligence Meets SQL Profiling

TigerGraph serves as the backbone for relationship mapping, allowing the agent to traverse connections between transactions, users, and devices. This is paired with DuckDB for rapid, local profiling of the raw data files. The result is an agent that doesn't just label a transaction as 'fraud' or 'not fraud,' but presents a structured argument with evidence from multiple angles.

Why Builders Should Care

For developers building AI agents, this shift from 'explanation after the fact' to 'investigation before the verdict' is a critical architectural upgrade. It moves the system closer to how human fraud analysts work. You aren't just trusting a black-box score; you're inspecting the graph traversal paths that led to the conclusion.

Key Takeaways

  • The project rejects the 'score-then-explain' pattern common in LLM fraud demos.
  • DuckDB was used for initial data profiling to inform the agent's logic.
  • TigerGraph handles the complex relationship mapping necessary for multi-sided arguments.
  • The demo includes a video walkthrough of the investigator in action.

The Bottom Line

Stop letting LLMs write fiction about your risk scores. Build agents that investigate, not just narrate.