ExceptionForge | Evidence Before Action

ExceptionForge is an evidence-first procurement investigation agent built for Vultr: Agent Rush 2026, Challenge 1, Blast Radius Zero. **The problem:** Purchase orders, goods receipts and supplier invoices often arrive in incompatible CSV formats, with inconsistent supplier names, quantities and column structures. Traditional matching can miss discrepancies or make investigations difficult to verify. **The solution:** ExceptionForge identifies potential procurement exceptions and provides source-linked evidence for human review. It does not authorise payments or execute financial transactions. **Public demonstration:** Hosted on Vultr Cloud Compute, the application processes 32 synthetic invoices through a clearly labelled deterministic analysis mode, identifies 15 exceptions and displays their source CSV row references. Users can inspect findings and export the results. **Real AI integration:** Separately, Vultr Serverless Inference generated dataset-specific Python, which ran inside a restricted Docker container. The recorded execution processed 32 invoices, producing 18 findings in 9.85 seconds. A real network-containment test returned "Network unreachable". The public demonstration and the live AI execution are separate, and the 18 AI findings have not been fully evaluated against ground truth. The Docker probe demonstrates a specific network restriction, not comprehensive isolation. **Technologies:** Vultr Cloud Compute, Vultr Serverless Inference, DeepSeek, Python, FastAPI, Docker and Nginx. **Video note:** The demonstration is subtitled but has no voice-over. Free-tier speech-generation restrictions, despite unused credits, prevented us from completing suitable narration before the deadline. **Core principle:** AI investigates. The sandbox constrains execution. Humans decide.