Inspiration

Scientists who study planetary surfaces switch between maps, datasets and papers all day, and AI tools make it easy to get confident answers that cite sources that don't exist. We wanted a spatial tool where you can point at a real lunar feature and get help you can trust, with the scientist always making the final call.

What it does

NEXUS is a WebXR app for Meta Quest. You place NASA's LRO Moon globe on a real table, open the Apollo 15 region as a 3D terrain model built from an LROC elevation model, and point at a feature such as Hadley Rille.

A team of 20 specialist AI agents then investigates it. A coordinator sends the work through seven phases: elevation and imagery, geology and history, competing explanations, contradictions and uncertainty, a draft, integrity advisors and next steps. The agents collaborate on a shared team board, and they can only cite evidence from a curated NASA evidence store.

Before anything reaches the headset, the Scientific Integrity Engine checks every claim. Evidence IDs must exist, numbers must match measurements computed by code, hypotheses must be worded as hypotheses, and contradictions must be shown. Jev, a decision model, weighs whether the evidence really supports a finding, and it can only make the gate stricter. Findings appear anchored to the terrain with their claim type, uncertainty and evidence. The scientist approves, rejects or marks each one inconclusive. No agent can mark a finding reviewed.

Everything works with hands alone: ray and pinch, fingertip poke and a palm-up wrist menu. Voice is only an optional shortcut.

How we built it

• Client: IWSDK and Three.js on WebXR, with hand tracking, passthrough and anchors. It is tested in the IWER Quest 3 emulator.
• Terrain: a deterministic measurement package, so depth, slope and distance come from code and never from a model.
• Integrity: a TypeScript Scientific Integrity Engine with validators V1 to V9 and a planted-bad-claim red-team suite that runs in CI.
• Agents: a 20-agent Docker Agent team defined in agents/nexus.yaml, with a read-only tool whitelist per agent. Jev decisions and Gemini reasoning are built behind adapters.
• Hosting: the public demo runs on Cloudflare Pages with no backend or keys in the client.

Challenges we ran into

• Making a trustworthy AI tool meant building the checking layer first, then letting the agents work inside it.
• The Quest frame budget limited the terrain mesh, so we built level-of-detail tiles from the elevation model.
• Testing hand interaction without a headset required driving the IWER emulator's hands.

Accomplishments that we're proud of

• Every claim passes an integrity gate, and the red-team suite must catch every planted bad claim.
• Jev can overrule a "sufficient" verdict only in the strict direction. In our live test, a planted claim that stated an interpretation as fact scored 0.02 for "supported" and was refused.
• A complete investigation works with hands alone.

What we learned

Trust is a design feature. Showing why a claim passed, what it cites and how uncertain it is matters as much as the answer.

What's built, and what's next (honest status)

• Public demo: it runs offline and replays research that was precomputed from archival NASA data and passed through the real integrity engine. Every card says so.
• Agent team: the 20-agent team runs live code in a development build (draft PR 15), but it is not deployed.
• Jev and Gemini: merged, and the Ask-a-question feature with live Gemini answers is on a preview site. The public demo uses the rule-based fallback.
• Not built yet: live speech recognition (speech is typed as simulated speech), spatial agent voices and Evidence Debate.
• Next: deploy the agent team behind a hosted backend, add speech input and spatial voices, and bring the same architecture to other environments such as oceans, archaeology and Mars

Built With

Share this project:

Updates

Submission history