Inspiration
How does fire behave in space? Almost all NASA combustion data comes from orbit (about 0 g) or from Earth (1 g). The Moon (0.165 g) and Mars (0.38 g) sit in between, with little or no direct data. Partial gravity is not bounded by the 0 g and 1 g results: some materials may burn more readily at partial gravity. Anyone planning a habitat needs to know where the evidence ends before relying on it. We built Oxyflame for the NASA Space Apps Challenge 2026 Flame in Freefall challenge to show exactly that.
What it does
- 7 crew environments (Earth, ISS, two Moon cases, a pressurized rover, a Mars habitat and a spacesuit), chosen once in a global scenario picker.
- Atlas ranks NASA investigations for the chosen scenario. An investigation that documents neither oxygen nor pressure shows "Cannot assess" instead of a made-up score.
- Ask answers in 6 parts: verdict, cited sentences, what differs, limit, next step and scope. Every number is checked against the records it cites. If a check fails, you get a template answer built directly from the records.
- Referee decides whether a difference between two runs can be pinned on one variable. It gives one of four verdicts: Can isolate, Cannot isolate, Can't confirm: missing values or No contrast between runs.
- Mission lab and Missing experiment show the nearest tested conditions, the evidence map, and proposed tests that would close the gap. These are illustrative, not an approved NASA protocol.
- From space to Earth links four everyday applications to their NASA sources: water-mist fire extinguishers, a smoke-vs-dust fire detector, firefighter air tanks and firefighter protective clothing.
- An installable mobile app that works offline, plus an SMS / WhatsApp webhook that replies using only the rules engine, with no LLM.
Verdicts rate how strong the evidence is. They never say a material is safe or unsafe.
Try it in 3 steps: pick FM2 habitat air (a Moon habitat, 0.165 g). On Ask, type "Would fabric burn differently in a lunar habitat?" and open a citation. Then open Atlas and Missing experiment.
How we built it
data/ curated NASA records: every value with its source, page and evidence type
packages/engine ranking, Referee, planner (deterministic TypeScript rules)
apps/worker Cloudflare Worker: retrieval, Gemini (wording only), citation validator, template fallback, cache
├─ apps/web website (React, Vite, Tailwind)
├─ apps/mobile installable mobile app
└─ /api/sms SMS / WhatsApp replies (engine only)
- Data: a Python ingest pipeline scrapes and normalizes NASA PSI records. Every value carries its source document, its page and an evidence type: Measured, Computed, Inferred, Assumed or Unknown. A missing value is Unknown with a reason, never 0 and never guessed.
- AI with guardrails: Gemini only proposes wording. It sees only the retrieved, ID-tagged records, and its answer is used only if it passes the schema check and every number matches a cited record.
- All on Cloudflare's free plan: Workers (API), Workers static assets (website), Workers AI
bge-m3embeddings, Vectorize (semantic fallback search), KV (answer and vector cache) and Pages (mobile app). If Workers AI or Vectorize is down, search falls back to keyword results.
Challenges we ran into
- Our own data was wrong at first. We set a rule that the person who collected a value never verifies it. The first independent check against the primary NASA PDFs confirmed only 3 of 13 runs and flagged 10. We removed a 101.3 kPa pressure that wasn't in the cited table and re-attributed two runs from BASS-II to BASS. A second pass then verified all 12 remaining runs.
- Semantic search lost. Hybrid search (keyword plus embeddings) scored Top-1 4/18 and Top-3 10/18, against 8/18 and 16/18 for keyword search plus the rules engine. So semantic search ships only as a fallback.
- SMS needs a real number. US carrier registration (A2P 10DLC) and WhatsApp sender verification are still pending, so the webhook is live but has no public phone number yet.
Accomplishments we're proud of
| Measure | Result |
|---|---|
| Runs verified against their NASA source | 12 of 12 |
| Engine-only ranking, Top-1 / Top-3 (18 labelled questions) | 9/18 / 15/18 |
| Referee: correct verdict / correct named reason | 9/9 / 5/5 |
| Adversarial traps caught (kPa vs psi, mole fraction vs partial pressure, burner vs cabin O2, liquid vs solid fuel, missing airflow) | 5/5 |
To be fair about these numbers: we adjusted the ranking rule against the same 18 questions, so the engine score shows fit to that set, not performance on new questions. The Referee pairs and traps were written by the engine's author.
What we learned
A "no data" answer is only useful if it is precise. Showing Cannot assess or Unknown (not documented) turned out to be more valuable than a confident guess. Independent verification also changed our dataset more than any code did.
What's next
- A fresh question set written by someone who hasn't seen the ranking rule, plus a second reviewer for the Referee tests
- Grounded-answer metrics against plain RAG (citation precision, unsupported claims, abstention)
- A live SMS / WhatsApp number once carrier registration clears
- More investigations beyond the current 25 investigations and 12 runs
Team
- Anish Shirodkar, AI lead: Ask pipeline, citation validator, Cloudflare Worker, website
- Navneet, data lead: NASA data collection, ingest pipeline, releases
- Juvana Khan, science and evaluation lead: engine, Referee, planner, verification, benchmark, mobile app, SMS
Scene photos are NASA public domain. AI coding assistants helped write code and docs, and Gemini writes the wording of Ask answers inside the app, constrained by the citation validator. Not affiliated with or endorsed by NASA. Sources: REFERENCES.md · Verification record: eval/verification.md
Built With
- cloudflare-kv
- cloudflare-pages
- cloudflare-workers
- css
- gemini
- html
- javascript
- python
- twilio
- typescript
- vectorize
- workers-ai