ExcursionLens

Inspiration

Cold-chain quality teams must reconcile temperature logs, custody events, vehicle telemetry, and stability guidance before deciding what deserves expert review. That evidence is fragmented, while an unconstrained AI can make dangerous release claims. ExcursionLens explores a safer pattern: deterministic measurements first, competing explanations second, and a mandatory qualified-human decision boundary.

What it does

ExcursionLens is an evidence-first investigation copilot for pharmaceutical cold-chain excursions. It:

• aligns multimodal sensor and custody events on one timeline;
• calculates breach duration, peak temperature, and degree-minute AUC deterministically;
• ranks competing causal hypotheses against supporting and contradictory evidence;
• runs an adversarial skeptic pass instead of accepting the first plausible story;
• provides a counterfactual workbench for testing preventive controls;
• generates an auditable quality-review packet with SHA-256 evidence digests;
• always ends in QUARANTINE_PENDING_QUALIFIED_REVIEW, never automatic product release.

How we built it

The backend uses Python 3.11, FastAPI, Pydantic v2, and deterministic domain engines. The operational interface uses React 18, TypeScript, and Vite. The inference boundary is architected for NVIDIA Nemotron through Nebius Token Factory with runtime model-catalog discovery and structured hypothesis/critique contracts.

This submitted demonstration runs in transparently labelled DEMO_FIXTURE mode because a live Nebius API credential was not available. No live Token Factory execution or served NVIDIA model is claimed. Fixture hypotheses are synthetic and pre-validated; deterministic telemetry calculations and the complete UI remain executable.

Challenges

• Preventing AI-generated safety or compliance claims from crossing the human-review boundary.
• Aligning asynchronous logger samples, custody events, door sensors, and power telemetry.
• Separating deterministic facts from hypotheses and counter-evidence.
• Producing auditable evidence while remaining honest when an external provider is unavailable.

Accomplishments

• 165 automated tests passed.
• 92.98% statement and 81.85% branch coverage.
• 12/12 frozen deterministic benchmark scenarios passed.
• 25/25 Playwright viewport executions passed.
• Automated Axe scans detected zero violations in the tested application states.
• Security scanning detected zero exposed secrets.
• The public build fails closed and clearly reports DEMO_FIXTURE / NOT_CONFIGURED instead of fabricating provider success.

What we learned

High-consequence AI needs bounded state machines, explicit abstention, contradictory evidence, and human authority. The most valuable agent behavior is sometimes refusing to conclude.

What's next

The next milestone is genuine Nebius Token Factory execution with an available NVIDIA Nemotron model, durable cloud hosting, authenticated evidence ingestion, organization-specific SOP adapters, and evaluation with qualified cold-chain professionals.

Important disclosure

This is investigation-support software demonstrated with synthetic or user-supplied evidence. It does not determine product safety, potency, regulatory compliance, or disposition. Follow organizational SOPs and obtain review from an authorized quality professional.

Runtime posture: INTERNAL_COMPLETE_SPONSOR_RUNTIME_UNVERIFIED. The public demo uses deterministic fixture replay; live Nebius/NVIDIA execution is not claimed.

Built With

• axe-core
• docker
• fastapi
• nebius-token-factory
• nvidia-nemotron
• playwright
• pydantic
• python
• react
• typescript
• vite