Ripple Room · Nebius × NVIDIA Global AI Hackathon 2026

See the chain reaction. Fix the day. Built at the Nebius × NVIDIA Global AI Hackathon 2026, track Best Apps and Agents.

When one thing goes wrong at a live event, Ripple Room shows everything it affects and helps you find a recovery plan you can trust. NVIDIA Nemotron on Nebius Token Factory proposes, an OR-Tools CP-SAT verifier checks, and a human approves.

• Live demo: https://ripple-room-five.vercel.app (no login). To control cost, live AI on the public demo may be paused; the badge in the top bar says which mode it is in. The solver, playbook search, preview, apply and undo always work.
• Built in October 2026 during the Nebius × NVIDIA Global AI Hackathon 2026 (28,000+ registered participants). It was not submitted for judging. Not affiliated with or endorsed by Nebius or NVIDIA.
• The event (Signal Night at Bay Hall), people and costs are fictional. "Ripple Room" is a working title, unrelated to Ripple Labs / XRP or other projects named "Ripple".

Who	Does	Cannot
NVIDIA Nemotron 3 Super on Nebius Token Factory	reads the situation through tools, proposes typed fixes, revises after a rejection	change the plan, or call a fix "verified"
OR-Tools CP-SAT + an independent validator	decides whether a fix holds, and explains why not	be overruled: a timeout is "unknown", never green
You	set priorities, compare verified options, approve one or none	nothing changes until you approve

Why it matters

Event coordinators run a day on one fixed opening time and many jobs that depend on each other. When a truck is late or a technician is out, they must work out quickly what else slips and what to change, usually from a spreadsheet that does not show the knock-on effects. Ripple Room shows the chain reaction, suggests fixes, and only offers a fix after a solver has checked that it really works. The coordinator stays in charge: nothing changes until they approve, and they get a list of who to tell.

Try it

1. Watch the day run: a venue scene driven by the plan's real times. Doors open at 18:00.
2. Press "What if the truck arrives 35 minutes late?" (or describe an incident in your own words). The delay ripples through the dependency map: doors would be ready at 18:15, 15 minutes late.
3. Set what matters (guest experience, spend, crew stability, budget, spare time) and press Find recovery options. Nemotron's proposals land on the verifier board, kept or rejected, each with a reason.
4. Preview an option on the whole day, approve it, and get the list of who needs to be told what, by when. Undo is one click.

How it works

Browser ──▶ FastAPI (stateless; client holds content-hashed revisions)
             ├─ impact: do-nothing propagation, causal waves, slack, driving path
             ├─ interpret: free text → typed incidents + priorities (Nemotron, one forced tool call) → human confirms
             ├─ recover (SSE): bounded agent loop ───▶ Nebius Token Factory ─▶ nvidia/nemotron-3-super-120b-a12b
             │     tools: get_situation · inspect_task · list_interventions · evaluate_patch · submit_options
             │     every proposal ─▶ verifier: patch checks → CP-SAT → independent validator → explanation
             ├─ playbook search (deterministic, labelled "no AI") and verified alternatives in the same pool
             └─ apply: re-verifies from scratch; stale revision → 409; change notice

• The model never edits the plan. It calls five allow-listed tools with strictly validated arguments. Only fixes the verifier accepts become options. A timeout is "unknown", never green.
• Rejections are explained from the solver's own results, for example "…makes things worse than doing nothing (18:15)".
• Priorities are soft, deadlines are hard. Options are ranked by priority breaches, then by a penalty breakdown shown on each card.
• Hybrid by design. Verified playbook options are always in the pool, so the final list is never worse than the playbook's best under the same ranking. Each option is labelled by origin (Nemotron, playbook, or no change).
• Raw model reasoning never leaves the server.

Claude Code was the development tool. At runtime the only AI is NVIDIA Nemotron served by Nebius Token Factory.

Measured results (live Nemotron 3 Super, 20 fixed cases, one run per version)

	Agent v2	Agent v4	Playbook search	Greedy
Verified recovery found when one exists	16/17	16/17	15/17	15/17
Honest "none" when none exists	3/3	3/3	3/3	3/3
Options failing independent re-validation	0	0	0	0
Best option equals the exhaustive oracle's	9/17	7/17	8/17	6/17
Median latency	19.5 s	18.4 s	0.08 s	0.11 s
Cost of the 20 runs (list price)	$0.24	$0.22	0	0

What this does and does not show:

• In case c19 (setup team away) the agent found a verified recovery that both deterministic baselines miss: a free person helps rig, plus a shorter line check. See docs/evaluation/CASES-c19-c09.md.
• It does not show better overall plan quality than the playbook. There is one run per version, and fixes that need three changes at once (c09) are still missed.
• Full data: docs/evaluation/. The current agent (hybrid, prompt v7) has not yet been measured live.

What is live, what is simulated, what is not done

• Live: recoveries and free-text incidents on the public demo call Nemotron 3 Super on Nebius Token Factory, unless the badge says the AI is paused or the day's budget is used. Solver, ranking, apply and notice are deterministic code.
• Simulated: the event, people and costs. The venue scene is illustrative (driven by the plan's times, not live tracking). Local runs and tests use a scripted mock agent labelled "Simulated agent".
• Not done: importing a real run-of-show, shared team sessions, sending the change notice. Fixes that need three changes at once (case c09) were found only by the exhaustive oracle.

Accessibility

An axe gate (WCAG 2.0/2.1/2.2 A and AA) runs on five journey states on desktop and mobile. Every control must be keyboard-reachable and named, and a keyboard-only journey test applies and undoes a recovery. Status is always written in words, not colour alone. Animations follow the system's reduced-motion setting, and the "Play animations" button under the venue scene opts back in.

Run locally

Prerequisites: Python 3.13 with uv, Node 24.

cd services/api && uv sync && uv run uvicorn app.main:app --port 8000          # mock agent by default
cd apps/web && npm ci && NEXT_PUBLIC_API_BASE=http://localhost:8000 npm run dev  # http://localhost:3000

Mock mode is labelled "simulated" in the UI. Live mode: copy .env.example to services/api/.env, set AI_MODE=live and NEBIUS_API_KEY, then uv run python ../../scripts/smoke_nebius.py --probe. Spend controls and the kill switch are in docs/operations.md.

Tests

cd services/api && uv run pytest -q && uv run ruff check . && uv run mypy app
cd apps/web && npm run typecheck && npm run lint && npm test && npm run build && npm run e2e   # desktop, mobile, Firefox
uv run python scripts/check_judge_demo.py https://ripple-room-five.vercel.app   # production judge path (makes live calls)

Repository map

services/api/app/domain scheduling domain · app/solver CP-SAT, verifier, playbook search, greedy · app/agent tools, prompts, loop, interpretation · app/providers Nebius adapter + labelled mock · app/evaluation 20-case harness · apps/web Next.js app · docs/ architecture, decisions, evaluation, operations, journal, feedback.

License

MIT. See LICENSE.