Varsha Shah cuts false compliance alerts by 76 percent using graphs

Most enterprise compliance systems check financial documents in a vacuum. A payroll registry matches internal rules, an invoice passes vendor validation, and a tax filing looks clean on paper. When audited independently, every transaction appears legitimate. However, modern corporate fraud hides in the white space between these isolated records. It exploits the blind spots of traditional Natural Language Processing (NLP) and rigid rule-based tools that cannot link records across disparate corporate networks.

At the AI Engineering World Fair, enterprise technical architect Varsha Shah introduced an AI-driven multi-document correlation framework designed to address this vulnerability. By moving away from document-level validation, this architecture introduces cross-document intelligence to discover hidden relationships across procurement, tax, payroll, and transaction systems.

The Three-Tier Cross-Document Architecture

The framework operates on three complementary layers that translate raw, disconnected enterprise data into structured risk intelligence:

• Entity Correlation Engine: A graph-based system that maps entities like employees, bank accounts, vendors, and transactions. It builds a unified relational graph, answering the fundamental question of what is connected.
• Adaptive Probabilistic Risk Model: Rather than relying on rigid, binary rules, this layer combines anomaly strength, historical audit patterns, and source reliability to calculate a confidence-based risk score.
• Cross-Jurisdictional Normalization: Global companies operate under diverse reporting standards. This component harmonizes currencies, tax rules, and local reporting periods so risk can be evaluated consistently across multiple international borders.

Testing the Framework Against Three Million Records

To prove the system's viability, Varsha Shah evaluated the model using approximately three million real-world financial records gathered over a five-year period across four regulatory jurisdictions.

The system achieved a 91% precision rate and an 87% recall rate, yielding a balanced F1 score of 0.89. More importantly for exhausted compliance departments, the framework delivered a 76% reduction in false positives and cut manual audit workloads by 40%. Because the system features a continuous feedback loop, verified fraud patterns reinforce future detection passes, while false alarms help refine and lower future risk-scoring thresholds.

Shifting to Predictive Corporate Governance

Moving beyond reactive audits allows companies to transition to predictive governance. Instead of conducting forensic autopsies to discover what went wrong months after the fact, compliance teams can monitor real-time data streams to flag anomalies before they trigger official regulatory violations. Success in deploying this system requires tight integration with existing enterprise resource planning (ERP) platforms, localized configuration for local laws, and direct alignment with active audit workflows to ensure investigators can act on high-priority alerts instantly.