SupplyGuard
- NVIDIA Nemotron
- Multi-Agent
- Nebius Token Factory
- Tavily
- FastAPI + SSE
- React
- Docker
Built at the Nebius Build Week Hackathon in partnership with NVIDIA ({Tech: Europe}, Berlin, 18 October 2026) for the Siemens Mobility: Supplier Risk Intelligence track, together with colleagues from ITUC — designed, built, and deployed to a Nebius VM in less than 6 hours.
The problem
Rail procurement teams juggle thousands of suppliers. Headline risk scores hide danger (911 of 10,000 suppliers rated Low/Medium carry a Critical sub-score), and signals — audit PDFs, supplier emails, news — are scattered across tools nobody links back to the score.
What we built
- A supervisor agent on NVIDIA Nemotron delegates in parallel to a Data Agent, Document Agent (Nemotron Parse OCR), Email Agent, and Web Agent (Tavily).
- Explainable risk, no invented numbers: we reverse-engineered the dataset’s 18-factor weighted risk index exactly — the LLM picks the tool and explains, the tool does the maths.
- Email-driven re-scoring: a supplier delay email triggers risk recomputation, line-stop analysis against safety stock, and alternative suppliers — applied only after human approval.
- Generated workspace: Nemotron writes a role-specific layout spec, validated against a fixed component list (the LLM never writes code).
- Security layer: prompt-injection guardrails, untrusted-content handling, user-defined boundaries, a Validator Agent, human-in-the-loop approvals, and an audit log.
Stack: Nemotron 3 Super, Nebius Token Factory & Cloud, Tavily, FastAPI with Server-Sent Events, pandas, PyMuPDF, React 18, Vite, Tailwind, d3, Docker, Caddy.