Noah Nvidia

What Noah does

Noah Nvidia is a supervised virtual employee for small service businesses. It turns an owner's plain-language business context into a structured, reviewable workspace, then turns everyday requests into proposed work that remains under human control.

How it works

An owner describes the business in natural language. The onboarding flow is designed to create a versioned onboarding.v1 JSON draft with business details, optional inventory, and missing fields. The owner can edit the draft before confirming it, or skip setup and explore the clearly labeled fictional Atlas Services workspace. Sensitive and external actions stay behind approval; the public synthetic sandbox does not send email, create calendar events, or move money.

Nebius and NVIDIA

The video shows a controlled test session with a real Nebius Token Factory call to NVIDIA Nemotron using fictional data in an isolated test tenant. It is not a recording of the shared public site's current runtime. As of September 23, 2026, the public demo is in deterministic synthetic mode and makes no model calls.

The public server-funded route is scheduled from October 27, 2026 at 17:00 UTC through December 16, 2026 at 00:00 UTC—four hours after judging ends on December 15 at 12:00 p.m. Pacific. This is bounded access, not full or unlimited access: the planned safety baseline caps it at 20 server-funded calls total and 5 per UTC day. The actual hosted opening depends on provider configuration/credit and durable Neon usage accounting; if credit, provider quota, configuration, or usage storage is unavailable, the demo remains or returns to synthetic mode. A separate reviewer BYOK route has its own limits of 5 total and 2 per UTC day per temporary key. The hosted settings and cutover still need verification.

OpenCode2API is a synthetic-only test transport and is disabled in the public configuration.

How I built it

I built the console with React and TypeScript, backed by a FastAPI service in Python. PostgreSQL on Neon stores tenant-scoped state and durable public usage reservations; Render hosts the demo. The onboarding schema is versioned, provider results are normalized, and approval actions are idempotent and auditable. Promptfoo evaluations and release checks cover extraction, provenance, safety boundaries, and fallback behavior.

What changed during the hackathon

I developed the product into a supervised onboarding and operations workflow: a versioned onboarding contract, natural-language extraction through the Nebius/Nemotron route, editable human review, explicit confirm-or-skip paths, tenant-safe persistence, bounded public usage, and reviewer guidance. I also hardened provider boundaries and made the public synthetic fallback explicit so simulated output is never represented as live inference.

Challenges and lessons

The central challenge was making model output useful without letting it silently become business state or an external action. I separated provider provenance from business data, made confirmation explicit, kept visitor credentials out of persisted data, and designed the public route to fail closed when its durable budget store is unavailable. The control loop is simple: describe, inspect, decide, then act under policy.