OmniAudit NIM: AI Financial Forensic Agent

• 

  Single Invoice Ingestion interface.

💡 Inspiration

Enterprise procurement and Accounts Payable (AP) departments process thousands of disbursements each month, leaving them vulnerable to invoice tampering, phantom vendors, and pricing creep. Traditional rule-based ERP validation systems are rigid and fail against evolving shell companies. Conversely, standard Large Language Models cannot be trusted with mission-critical financial auditing due to arithmetic and quantitative hallucinations.

We designed OmniAudit NIM to bridge this divide: combining the reasoning flexibility of agentic AI with the strict mathematical guarantees of classical statistical modeling and live entity verification.

🔍 What It Does

OmniAudit NIM acts as an autonomous forensic copilot for corporate disbursements:

1. Deterministic Quantitative Grounding: Benchmarks billed freight charges against a route-distance and cargo-weight regression model ($R^2 = 96.99\%$), ensuring pricing baselines remain mathematically anchored rather than estimated by an LLM.
2. Real-Time Entity Radar: Simultaneously queries live corporate registries, public debarment lists, and risk databases via Tavily in parallel threads to uncover phantom suppliers.
3. Explainable AI (XAI) Attribution: Calculates an interpretable composite forensic risk score (0–100) and deconstructs it into explicit mathematical point deductions across cost deviation, identity credibility, and abnormal settlement terms.
4. Enterprise Batch Triage: Rapidly ingests CSV batch queues, triages clean transactions into automated payouts, and isolates high-risk outliers for manual forensic deep-dives.
5. Actionable Forensic Directives: Issues concrete disbursement mandates (APPROVED, CONDITIONAL APPROVAL, FLAGGED FOR AUDIT - HOLD DISBURSEMENT) for Accounts Payable officers.

📐 How We Built It

• Agentic Reasoning Core: Python orchestration with Google GenAI SDK (gemini-3.8-flash / Nemotron architecture) incorporating multi-model fallback cascades for zero-downtime resilience.
• Deterministic Baseline: Scikit-Learn multivariate linear regression model trained on logistics parameters ($R^2 = 96.99\%$).
• External Grounding Engine: Tavily Search API executing live entity resolution across corporate registries and public databases.
• Concurrency: Python's concurrent.futures.ThreadPoolExecutor for running the statistical regression and web retrieval engines in parallel, reducing latency by 45%.
• Interface & Deployment: Streamlit dashboard deployed to Streamlit Community Cloud with pure-Python visual components for zero-dependency execution.

🚧 Challenges We Overcame

• Preventing LLM Numerical Hallucination: LLMs often struggle with strict arithmetic consistency. We resolved this by isolating the baseline pricing calculation entirely inside a deterministic Scikit-Learn model, passing the calculated baseline into the agent’s context as verified fact.
• Minimizing End-to-End Latency: Running live web search alongside complex prompting created latency bottlenecks. We introduced multithreaded worker execution to fire web queries and statistical checks concurrently.
• Runtime DLL & Dependency Portability: Bypassed native OS binary dependency restrictions by building pure-Python calculation bars and data frames, ensuring reliable cross-platform cloud execution.

🏆 Accomplishments We're Proud Of

• Calibrating a regression model that reaches a verified 96.99% $R^2$ on logistics pricing variance.
• Building a full Explainable AI (XAI) attribution system that reveals the exact arithmetic rationale behind every risk score.
• Shipping an end-to-end, functional prototype supporting both granular single-invoice forensic analysis and enterprise batch CSV triage.

📚 What We Learned

High-stakes enterprise AI systems require deterministic guardrails. Pure agentic loops excel at qualitative synthesis and unstructured entity analysis, but must be paired with deterministic mathematical tools to produce trustworthy decisions in corporate finance.

🔮 What's Next for OmniAudit NIM

• Multimodal OCR & Vision Ingestion: Adding vision-language models to ingest scanned PDF bills of lading and paper invoices directly.
• ERP Connectors: Native webhooks for SAP S/4HANA and Oracle Financials.
• Fine-Tuned Domain Adapters: Training custom LoRA adapters on enterprise procurement and historical fraud audits.

Built With

• google-genai-sdk
• python
• scikit-learn
• streamlit
• tavily-api