Crew Air

Smoke and heat forecasts turned into a shift plan, so the job site keeps working and stays within the law.

Stanford Cleantech Hackathon 2026 · submitted to Track 3 Resilient Systems (line crews after a fire; the same hours for a home and a building, airsignal.app/crew/grid) · Live: airsignal.app/crew · Devpost · Deck (PDF) · PPTX · Explainer video · pitch/ · Write-up on rustman.org

The problem

California already writes down what an outdoor employer must do: at PM2.5 AQI 151 provide N95s and tell the crew (Cal/OSHA §5141.1), at 80°F shade and water, at 95°F high-heat procedures (§3395). Almost nobody does it.

• In the 2020–21 fire seasons California farmworkers spent at least 4.5 million worker-days at AQI 151+ (our calculation from EPA AirData × BLS QCEW, docs/evidence.md). Cal/OSHA issued 11 smoke-rule citations from July 2019 to October 2021.
• About 60% of 300+ Central Valley farmworkers got no N95 from their employer (CCEJN survey).
• Heat is the opposite: the most-cited Cal/OSHA standard in agriculture every year, and the classic miss is no written plan in the field.

Inspection waits for complaints, and workers are afraid to complain.

What it does

1. Shift plan. Place, day and shift in; the hour each rule starts, the action it requires, the clear window for heavy work, and when a building should close outside air (BMS). One printed sheet with a pre-shift briefing sign-off.
2. Checked four ways. Model (hourly forecast, AQI as an EPA NowCast, not the 24-hour mean forecast feeds give), satellite (NASA VIIRS fires within 50 km → warning), ground (AirNow monitors within 25 km; a 1.5x disagreement scales the forecast toward the monitor, never below half), and the crew.
3. Check the sky. A photo of the horizon (phone or the page) goes to TypeSafe's Jev, a vision decision model that answers typed questions with probabilities. The phone blurs faces and crops to the sky first. On 12 public-domain photos with known answers it made the smoke call right 11–12/12 across runs, never wrongly (work/sky_eval.py, work/sky_eval.md); its visibility estimate was unreliable, so the verdict rests on the smoke question and visibility shows only when confident. Visible smoke under a forecast below AQI 101 is flagged as "looks worse than the forecast" and pre-fills a report.
4. Worker reports, the ladder. A worker reports smoke, ash, symptoms, or missing N95s, shade or water. No name, and the evidence for that hour attached. The crew lead marks it handled. After 30 minutes an open report can go further, anonymously (Cal/OSHA complaint link is real; union and legal-aid forwarding is a mock until they agree).

curl 'https://airsignal.app/api/crew?place=Fresno&shift=07-17'          # plan + checks + BMS, home, alerts, wind, fires_near, dispatch (JSON)
# every endpoint: https://airsignal.app/api · a utility's 13 cities: https://airsignal.app/crew/grid
curl 'https://airsignal.app/api/crew/reports?place=Fresno'               # today's reports
curl -H 'content-type: image/jpeg' --data-binary @sky.jpg \
     'https://airsignal.app/api/sky-check?place=Fresno'                  # Jev on a sky photo
make plan P="Fresno" S=07-17                                             # the same engine idea in Python
make test                                                                # Python rules + NowCast tests
make evidence                                                            # the 4.5M worker-days table

Where the code is

Piece	Where	Built during the hackathon
Engine: NowCast, rulebook, plan, BMS, fire and monitor checks	airq airq-core/src/crew.rs (Rust → WASM)	yes, all of it
Page, API, reports, escalation mock, sky check	air-signal-web src/pages/crew.astro, src/lib/crew.ts, src/pages/api/crew/, src/pages/api/sky-check.ts	yes, all of it
Python CLI + rulebook YAML + tests, evidence script	this repo: crewair/, rulebooks/, tests/, work/	yes
iOS app: Plan, Sky (AR horizon + Jev check), Report, crew lead	this repo: ios/ (make -C ios gen test build install)	yes
Evidence and the Jev eval	this repo: work/evidence_*.py, work/sky_eval.py	yes
Pitch	pitch/	yes

Every commit for the above is dated 23–24 Oct 2026 inside hacking hours (Fri 17:00 – Sat 17:00 PT).

After the event the iOS code continues in air-signal-ios (private, Crew Air plus a consumer app on a shared core); ios/ here stays as built in the 24 hours.

Earlier open-source work it stands on, disclosed: the Air Signal site's hosting and layout (airsignal.app, air-signal-web), and from the airq-core library the EPA AQI breakpoints, the sensor/model merge (merge.rs) and haversine. The iOS AR overlay adapts camera-geometry code from our app Starlapse. The app's usage counter is our shared package superduperai-analytics-swift (three events, an install UUID, August 2026). All four are public repositories by the same author, older than the event. The rest is new.

Rulebooks: one file per jurisdiction

The engine knows no law. A rulebook names a metric (us_aqi_pm25, pm25_nowcast_ugm3, pm25_ugm3, temp_f, temp_c), a threshold, the action and the citation. Smoke rules use the NowCast concentration as §5141.1 Appendix A does (55.5 and 500.5 µg/m³): an AQI threshold of 501 can never trip, the scale stops at 500. A new state or country is a new file, not new code: Oregon and Washington have smoke rules, Canadian provinces use AQHI, Spain and Italy restrict outdoor work in heatwaves. Only California ships.

Data sources

Source	Key	Use
Open-Meteo (CAMS air quality + weather)	none	hourly PM2.5 and temperature; free only non-commercially
NASA FIRMS VIIRS, public 24 h CSV (NOAA-20, Suomi-NPP)	none	fires within 50 km
AirNow hourly files (files.airnowtech.org)	none	regulatory monitors within 25 km
EPA AirData, BLS QCEW	none	the evidence table
DOL OSHA enforcement API	free key	§5141.1 / §3395 / §3396 citations (work/evidence_citations.py)
TypeSafe Jev	key (Worker secret)	sky photo → smoke / visibility probabilities
Earth Engine, Sentinel-2	gcloud login	burn scar and fire front for the pitch (work/)
NOAA HRRR-Smoke	none	next: 3 km smoke over the US instead of CAMS ~40 km

Model

Free for crews, paid by employers, public data. The engine, rulebooks and worker app are open source (AGPL for airq-core and the site, MIT here). Employers pay per site for alerts in Spanish and English, many sites, the archive an inspector or insurer asks for, and the BMS feed. Raw reports stay private; exposure and report counts by county are published anonymized.

Satellite side (Earth Engine)

work/ holds the GEE 101 workshop run locally plus two scripts on the largest fire in California on 20 Oct 2026 (NASA FIRMS, near Sequoia): fire_s2.py pulls a Sentinel-2 scene, burn_scar.py computes dNBR. Auth: ee_init.py uses your gcloud login (EE_ACCOUNT, EE_PROJECT).

Plan and status: docs/next.md. Event notes: docs/event.md, tracks and decision: docs/tracks.md.

A planning aid, not legal advice.

License

MIT