EquatorGuard: technical documentation, assumptions and limitations
EquatorGuard is a heat-risk early warning tool for outdoor crews. It answers what a site-wide reading cannot: which workers face the most risk today, and what should the supervisor do?
Working prototype for the Future Health Digital Innovation Challenge (Population Health · Climate-Resilient Worker Health). Real weather data, synthetic demo roster. Not yet piloted. Not a medical device.
1. How it works
Built for Abu Dhabi first. Abu Dhabi's code of practice (ADOSH-SF CoP 11.0) manages heat with Thermal Work Limit zones measured on site. TWL assumes workers are acclimatized, and its zone table says "No unacclimatised person to work" in Medium and High Risk zones: new workers, or anyone off work more than 14 days. A heat meter cannot know who that is. An attendance system does. EquatorGuard joins the two.
- The supervisor enters the site's TWL zone from the instrument. EquatorGuard shows the official controls for that zone, word for word from Table 1, and lists who must not be assigned.
- From attendance (demo roster, or a CSV the user loads, processed in the browser only): new starters, days away, long runs of shifts. Each worker gets today's allowed share of work from the guideline's acclimatization and re-acclimatization tables. No medical data is needed.
- An hourly heat-load forecast (estimated WBGT from weather data) ranks the day and finds the coolest window for heavy tasks.
- Crew instructions and individual decisions are acknowledged by a supervisor and kept in an audit trail. Worker messages are ready in English, Arabic, Hindi, Urdu and Indonesian. If signs of heat illness are reported, the guideline's first-aid steps are shown and an escalation is logged.
Where TWL is not used (the second demo site, an open-pit coal mine in East Kalimantan, Indonesia, with a synthetic crew), the same engine applies NIOSH limits to estimated WBGT and gives allowable work minutes per hour.
Rules decide. AI explains. A person acts. Levels and actions come from a deterministic rule engine. A language model only words the shift briefing, from facts that are shown next to it.
2. Architecture
| Part | What it does | Technology |
|---|---|---|
| Rule engine | TWL zone rules (Table 1), acclimatization tables, NIOSH limits and allowable minutes, levels, crew summaries, CSV import, first-aid strings | TypeScript, pure functions, unit-tested. Runs in the browser in the prototype; on the server in a pilot, so decisions are logged at the source. |
| Worker API | Fetches the hourly forecast and computes estimated WBGT for sun and shade; stores the audit log and the last good forecast; calls the language model for the briefing | Cloudflare Workers, Cloudflare D1, Cloudflare Workers AI (Meta Llama 3.3 70B Instruct) |
| Dashboard | Supervisor and site-manager views, worker drawer, site display, measures, back-test, data and governance pages | Plain TypeScript, no UI framework, hand-written SVG charts; about 41 kB compressed |
| Offline layer | Keeps the app and the last forecast on the device; says when it is showing saved data | Service worker, network-first pages, cached hashed assets |
Data flow: site instrument reading (TWL zone or WBGT) + attendance and roster + hourly weather → rule engine → crew instructions, individual decisions, coolest window, worker messages, site display → supervisor acknowledges → audit trail. Fallback: weather service unreachable → last good forecast → manual entry of the site reading.
3. Methods
Estimated WBGT
Liljegren et al. (2008), as implemented in ECMWF thermofeel 2.3.0 (Apache-2.0) and used operationally by KNMI, ported to TypeScript. Inputs from Open-Meteo: air temperature, relative humidity, surface pressure, 10 m wind, instantaneous global and direct radiation, and the solar zenith angle (NOAA solar position). Covered or indoor work uses the same method with solar radiation set to zero. The result is an estimate for an unshaded point from gridded weather, not an on-site reading.
NIOSH limits and allowable work minutes
NIOSH (2016) RAL = 59.9 − 14.1 log₁₀M and REL = 56.7 − 11.5 log₁₀M (°C-WBGT, M in watts; §8.1). The REL applies to acclimatized workers and the RAL to unacclimatized workers. Metabolic heat is taken on a 1-hour time-weighted basis including rest (§1.1.3). Workload classes: light 200 W, moderate 300 W, heavy 415 W, very heavy 520 W, each inside its NIOSH class (Table 5-1); resting 117 W. Clothing adjustments from Table 3-2 are added to the WBGT. Allowable work minutes = the share of the hour at the task rate, with the rest at 117 W in the same conditions, that keeps the 1-hour average within the worker's limit.
Abu Dhabi zones and acclimatization
ADOSH-SF Technical Guideline – Safety in the Heat, v4.0 (15 July 2024): Table 1 (zones, interventions, rehydration and work-rest schedules, shown verbatim), §3 (acclimatization: new employee 20/40/60/80/100%, experienced employee 50/60/80/100/100%; re-acclimatization by days away), §4 (signs and first-aid steps, shown verbatim). Unacclimatised = new workers, or off work more than 14 days (Table 1, note b).
Levels
Abu Dhabi site: the level follows the entered zone and Table 1 (work allowed with a work-rest cycle is "Restricted"; people who may not be assigned are listed separately). NIOSH site: level cut-offs on allowable minutes per hour at 60, 45 and 15 (configurable site rule).
Configurable site rules (not NIOSH or ADOSH values)
- Acclimatized after 7 workdays in heat (NIOSH: "over 7–14 days"); absence reset after 7 calendar days away (NIOSH site).
- Cumulative load: 7 or more consecutive shifts (no weekly rest day) or a shift longer than 10 h. On NIOSH sites it raises the level one step only when heat is already a factor; on ADOSH sites it triggers an individual review and never changes the Table 1 level.
- A new worker is one in the first 7 workdays (CoP 11.0 §3.2(b)(i) allows 5–7 days).
- Moderate, heavy and very heavy tasks all map to Table 1 "Heavy work".
4. Data sources
| Source | Type | Owner | Personal data |
|---|---|---|---|
| Weather forecast and history | Environmental, hourly, gridded | Open-Meteo (forecast; historical best match of ECMWF IFS, ERA5, ERA5-Land); CC BY 4.0 | None |
| Attendance and shifts | Operational | Employer; in the prototype synthetic, or a CSV the user loads, processed in the browser only | Pilot: worker ID, shift times, days worked. Prototype: none stored |
| Task and PPE profile | Operational | Site supervisor | None |
| Site instrument reading (TWL zone or WBGT) | Environmental measurement | Site HSE team | None |
| Acknowledgement log | Operational record | Site | Worker ID and action taken. Escalations are logged at crew level, without the worker or the symptoms. |
Not collected, by design: health status, medical history, wearable or biometric signals, location tracking, productivity figures.
5. Assumptions
- Hourly gridded weather for the nearest grid cell represents conditions around the site closely enough to rank hours and to plan; it does not replace the site instrument.
- NIOSH limits apply to a "standard" worker of 70 kg and 1.8 m² body surface wearing conventional work clothes, adjusted for PPE with the NIOSH clothing factors.
- Each task maps to one fixed metabolic rate (200, 300, 415 or 520 W); rest is taken at 117 W in the same conditions.
- An attendance row means the worker was on site that day. A worker with no row for the assessed day is treated as not on shift; days away count up to the assessed day.
- The TWL zone entered by the supervisor reflects the site instrument for that hour and applies until it is changed.
- The UAE midday break (12:30–15:00, 15 June – 15 September) applies to sun-exposed work only.
- The site follows the published ADOSH-SF tables; the prototype uses the absence column of the re-acclimatization table and, where 12 days appears in two printed rows, the stricter row.
6. Limitations
- It is an exposure estimate, not a health assessment: it cannot see how an individual worker is coping.
- EquatorGuard does not compute TWL. A pilot would add the published TWL calculation and validate it against the site instrument.
- Gridded weather is not an on-site reading; shade, reflective surfaces and machinery heat change conditions on site. The NIOSH limit assumes a continuous hour of work.
- Attendance cannot show where leave was spent; whether leave was in a cool climate is confirmed by the supervisor with the worker. Illness days are not in attendance data.
- TWL zone entries are kept on the device where they were entered; a pilot shares them between the supervisor's phone and the site display.
- Nothing is sent to workers automatically; messages are fixed templates, and every translation except English is a draft pending native review.
- The language model can word a briefing badly; it is told to use only the facts given, and the facts are shown. If it is unavailable, template text is used.
- Alert lead time in a replay is an upper bound, because observed weather stands in for the forecast.
- Integration with a real attendance system is by CSV only in the prototype. No field validation has been done and no effect on health outcomes has been shown.
- Weather data uses Open-Meteo's free tier, which is for non-commercial use; a paid pilot needs a commercial plan, national data or site sensors.
7. Validation
- The WBGT port agrees with ECMWF thermofeel within 0.05 °C on all 2,232 hours of Abu Dhabi summer 2026 (
tests/parity.test.ts). - 34 automated tests cover NIOSH limits, attendance and acclimatization rules, the Table 1 mapping, crew summaries, the empty state with no zone, CSV import edge cases and the first-aid content.
- A separate AI code review was run on 16 November 2026 and its findings fixed, including an absence-gap bug that could have hidden a returning worker, and offline precaching from the first visit.
- Offline use was tested on the deployed version: one online visit, then no network; the app loads and shows the saved forecast with an "Offline" notice.
- Summer 2026 back-test for Abu Dhabi (reproducible from the raw weather file): estimated WBGT was above the NIOSH limit for continuous moderate work in 96% of daytime hours of the midday-break season, and in 91% with the sun ignored. In 2026 the Abu Dhabi grid cell reached 35 °C on 184 days, 91 of them outside the midday-break season.
8. Privacy, security and responsible AI
- Worker protection first: no productivity figures, no ranking of workers, no disciplinary or employment use.
- Data minimisation: no health data; reported signs are used only to show first-aid steps and are not stored.
- Role-based views: site managers see crew totals only; supervisors see their crew because they assign the work.
- Human in the loop: the rules decide, the model only explains, and a person acknowledges and logs every action.
- Security: content security policy and related headers; API inputs validated and length-limited; the AI endpoint is cached and capped per day; the demo log is capped and cleaned up after 14 days. No cookies or analytics.
- Residency: a UAE pilot that handled health data would need in-country hosting or an approved exception; this version holds no health data.
9. Deployment and scale
Serverless on Cloudflare's edge network, with no servers to run. A new site is a configuration entry (location, time zone, site standard, crews). The page is about 41 kB compressed and works on any phone, tablet or TV a site already has; there are no sensors or wearables to charge or to fail in dust. Path to pilot: one Abu Dhabi site with a contractor's HSE team and an occupational health partner; attendance by CSV export first, then API; TWL readings entered by hand first, then read from the instrument's export; the rule engine moves to the server so decisions are logged at the source; in-country hosting.
10. API
| Endpoint | Purpose |
|---|---|
GET /api/weather?site=abudhabi|ekalimantan | 48 hours of forecast with estimated WBGT for sun and shade; falls back to the last good forecast |
GET/POST/DELETE /api/acks?session=… | Demo audit trail, scoped to an anonymous session; capped per session and per day |
POST /api/briefing | Shift briefing worded by the language model from the facts given; template text if unavailable |
11. References
- ADPHC. ADOSH-SF Code of Practice CoP 11.0 – Safety in the Heat, v4.0, 15 July 2024; ADOSH-SF Technical Guideline – Safety in the Heat, v4.0, 15 July 2024.
- NIOSH (2016). Criteria for a Recommended Standard: Occupational Exposure to Heat and Hot Environments. DHHS (NIOSH) Publication 2016-106.
- Liljegren JC et al. (2008). Modeling the wet bulb globe temperature using standard meteorological measurements. J Occup Environ Hyg 5(10):645–655. ECMWF thermofeel 2.3.0.
- OSHA. Protecting New Workers. osha.gov/heat-exposure/protecting-new-workers.
- Flouris AD et al. (2018). Lancet Planet Health 2(12):e521–e531.
- Weather data by Open-Meteo.com (CC BY 4.0). Generated using Copernicus Climate Change Service information 2026.