MamaCOCO

INQUEST - One Alert. Multiple Perspectives. Verified Evidence.

Inspiration

In gaming and lottery environments, detecting unusual financial activity is only the first step. Investigators must determine whether an alert reflects a genuine concern or behavior that has a reasonable explanation.

Traditional rule-based monitoring can identify anomalies, but interpreting those anomalies requires reviewing transaction histories, checking context, and documenting evidence. This can be time-consuming, especially when many alerts compete for attention.

We built INQUEST for the ThunderHacks II OLG Intelligent Player Protection and Fraud Detection Challenge to bridge the gap between automated detection and responsible human investigation.

Our central idea is simple: an AI investigation should challenge an alert, not merely confirm it.

What it does

INQUEST is an AI-assisted investigation platform that helps analysts detect, prioritize, investigate, and review potentially suspicious gaming transactions.

The platform combines:

Explainable risk scoring: Seven predefined behavioral rules analyze transaction frequency, transaction amounts, unusual locations or low play-through, transaction timing, historical spending changes, and repeated transaction patterns.
Contextual alerts: Each subject's recent seven-day activity is compared against their previous 23 days of transaction history.
Investigation prioritization: Cases are organized by risk score, the number of triggered indicators, urgency, and evidence completeness.
Two opposing AI investigators: A Suspicion Agent searches for evidence supporting concern, while an Alternative Explanation Agent examines benign explanations, contradictory evidence, and missing context.
Evidence validation: Deterministic checks verify supporting evidence references, numerical claims, and selected inconsistencies before findings are consolidated.
Lead AI assessment: The Lead AI synthesizes the investigation into a provisional assessment without automatically declaring fraud.
Human-in-the-loop review: Investigators retain control over case decisions, with additional senior review for substantiated findings and recorded review history.

For example, a sudden increase in spending may generate an alert. Rather than automatically treating that increase as fraud, INQUEST examines transaction frequency, amounts, spending history, repeated patterns, and available contextual evidence.

The goal is to help investigators make informed decisions rather than replace their judgment.

How we built it

We developed INQUEST as a full-stack web application using React, TypeScript, Vite, Python, and FastAPI.

The backend processes synthetic lottery and online gaming transaction records. A deterministic detection engine calculates a risk score from transparent, weighted behavioral indicators. A supplementary Isolation Forest anomaly model provides another analytical signal, but does not contribute to the operational risk score.

For AI investigations, we integrated the OpenAI API with two specialized agents that independently investigate opposing interpretations of a case. The agents access case-specific evidence through read-only tools rather than receiving unrestricted access to the database.

A deterministic validation layer checks evidence-backed findings before a Lead AI synthesizes the results. The system also supports saving investigation runs and recording human review outcomes.

We use PostgreSQL through Supabase for persistent storage, FastAPI for backend services, and React for the investigator interface. Our deployment architecture uses Render and Vercel.

All demonstration cases use synthetic data. No real OLG customer records are included.

Challenges we ran into

Balancing fraud detection with false positives. Suspicious behavior is not necessarily fraudulent. We addressed this by separating behavioral scoring from investigation and introducing agents that deliberately consider competing explanations.

Keeping AI findings grounded in evidence. Large language models can make unsupported claims or misinterpret numbers. We introduced case-scoped evidence tools, citation requirements, and deterministic validation checks to reduce these errors.

Handling contradictory evidence. During testing, we found situations where contextual events could be overused to explain away suspicious activity. We strengthened our checks to examine whether an explanation actually accounts for the observed transactions.

Building a complete investigation workflow. We worked on connecting alert generation, case prioritization, AI analysis, saved reports, and human review into one usable interface.

Deployment and database integration. Moving from a local application to PostgreSQL and cloud hosting required resolving connection-string formatting, encrypted connections, and deployment configuration.

Accomplishments that we're proud of

We are proud of building a working prototype that goes beyond simply labeling transactions as suspicious.

Our key accomplishment is a structured investigation workflow in which competing AI agents examine the same case from different perspectives, their findings are checked against evidence, and the final decision remains with a human investigator.

We also implemented transparent risk scoring, contextual case prioritization, persistent investigation history, and an investigator-focused interface.

The system demonstrates how explainable detection and AI-assisted reasoning can complement each other in a sensitive domain.

What we learned

We learned that detecting an anomaly and investigating it are fundamentally different problems.

We also learned that using multiple AI agents does not automatically guarantee correctness. Their outputs need structured tools, evidence constraints, validation, and human oversight.

Working on INQUEST reinforced the importance of explainable systems, careful handling of incomplete data, automated testing, and designing AI around real investigative workflows.

What's next for INQUEST

The next stage would be testing against more realistic, independently labeled datasets and working with domain experts to assess false positives, investigator efficiency, and decision quality.

Future improvements could include stronger payment and device-level evidence, more advanced network analysis, configurable institutional risk rules, stronger production authentication and authorization, and additional evaluation of how AI-assisted investigation affects reviewer decisions.

Our long-term goal is to turn INQUEST into a trustworthy investigation assistant that helps analysts work more efficiently while keeping people responsible for consequential decisions.## Inspiration