Onus HybridMind: AI Procurement Auditor

Enterprise procurement teams are bound by complex vendor contracts containing rebate thresholds, volume discounts, and penalty clauses buried in dense legal documents. Meanwhile, SQL databases hold the transaction records, but no one cross-references them against contract terms at scale. The result is millions in uncollected rebates and compliance gaps going undetected. To solve this, we built HybridMind: an autonomous AI audit agent that bridges structured SQL procurement data and unstructured vendor contracts in real time. It is not a standard RAG chatbot; it is a deterministic, multi-agent auditor built specifically to stop financial leakage. Our backend leverages LlamaIndex Workflows to orchestrate three distinct AI agents powered by Gemini 3.1 Flash Lite. The process begins with the Executor Agent. It converts natural language into strict SQL queries against our Supabase PostgreSQL database while simultaneously querying our ChromaDB vector store for the exact legal clauses. Next, the Verifier Agent takes both data sources and performs logical validation. It cross-references the SQL math against the PDF rules to detect discrepancies, such as verifying if a 12,000 unit purchase correctly triggers a 10,000 unit rebate threshold. Finally, the Chronicler Agent packages the verified finding and broadcasts the financial metrics to our live React dashboard via WebSockets. This architecture ensures transparent reasoning. While standard AI often hallucinates numbers, HybridMind forces the AI to show its raw SQL, explicitly cite its contract clauses, and prove its math. By keeping the frontend stateless and the backend firmly grounded in dual-silo retrieval, nothing is hidden behind a black box. Ultimately, HybridMind turns silent leakage into documented liability.