Inspiration

In a Central Valley summer, high heat and wildfire smoke can land on the same day, and a crew lead has to pick tomorrow's start time tonight. Cal/OSHA rules switch on at thresholds (80°F shade, 95°F high-heat procedures, AQI 151 for the wildfire smoke rule), but a threshold only tells you when you've crossed a line. It doesn't help you rebuild the shift. Madrugada is Spanish for the hours before dawn, and we wanted to know: does moving work earlier actually reduce time in the worst heat, and by how much?

What it does

Madrugada is a day briefing for crew leads, foremen and farm labor contractors:

Today: first light, the best work plan that starts no earlier than first light, and forecast temperature and air quality against the Cal/OSHA trigger lines, for a recorded real day (every day carries a REAL or SYNTHETIC badge, and a REPLAY label).
Planner: paid hours are an input the tool holds constant. It compares our plan with a fixed 7:00 shift, a fixed 5:30 shift and a heat-only planner, in modeled exposure-hours.
The hours scheduling can't fix: the high-heat hours left even in the best plan from first light. On the default day that's 3.5 modeled hours at or above 95°F. Those hours need shade, rest or crew changes, not a different start time.
Crew sheet: a one-page, English and Spanish printable sheet with the plan and that line.
Evidence: our preregistered results, each with its own confidence interval.
A narrated opening sequence (with a presenter mode) tells the story with numbers read live from the result files.

It's an advisory aid in modeled exposure-hours. It isn't a compliance tool or a health prediction, it doesn't replace the employer's written Cal/OSHA Heat Illness Prevention Plan, and it tracks no individual workers.

How we built it

Backend: Python, FastAPI and numpy. The day is a grid of 96 fifteen-minute slots. Each slot gets a cost from config-file thresholds:

$$ c_s = w_h\,h(T_s) + w_m\,m(\mathrm{AQI}_s) + w_d\,\mathbf{1}[\text{dual hazard}] $$

and the optimizer does an exhaustive search over every shift of one or two blocks using prefix sums. No solver and no machine learning, so every answer can be checked.

Forecast without peeking: the plan for day $ D $ uses only earlier days, $ \hat{x}D = 0.5\,x{D-1} + 0.5\,\bar{x}_{D-7..D-1} $. An ORACLE mode that uses the day's own data is labeled as an upper bound everywhere.
Real data, no API keys: IEM ASOS airport weather and EPA AirData PM2.5, converted to local clock time. PurpleAir readings can be corrected with the EPA US-wide formula $ 0.524\,\mathrm{PM}_{cf1} - 0.0862\,\mathrm{RH} + 5.75 $.
Evaluation: two windows of real summers, 1,296 region-days (2022–2025) and 768 region-days (2018–2021). Each was preregistered with SHA-256 hashes of the code and config before it ran. Confidence intervals use a bootstrap that resamples whole dates, and the two windows are never pooled.
Frontend: one offline HTML file with vanilla JavaScript, inline SVG charts and a canvas opening sequence. It has light and dark themes, works from phone to desktop, and needs no build step or CDN.
Guardrails: 184 automated tests, including a claims audit that fails the build if our own text overclaims.

Challenges we ran into

The data said no on smoke. We started out building a dual heat-and-smoke optimizer. On real data it showed no smoke benefit over heat-only planning, so we repositioned it as a heat-first planner that shows smoke, and said so.
The benefit depends on the start time. With a 4:00 AM earliest start, modeled hours at or above 95°F fell by 0.38 h/day versus a fixed 5:30 shift (window 1 CI −0.43 to −0.33; window 2 −0.38, CI −0.44 to −0.32). With a 5:30 earliest start it disappears (window 1 +0.004 h/day, CI +0.001 to +0.008). That changed the product: it now leads with the hours scheduling can't fix.
Getting real data without keys: time zones and daylight time, monitors several kilometres from the weather stations, and missing days.
Honesty under time pressure: we kept every number tied to a result file and wrote a test that blocks words like "guarantees".

Accomplishments that we're proud of

Two preregistered real-data evaluations, with the result replicated in both windows.
An honest product turn: the main readout is what scheduling can't fix.
A polished offline app with a bilingual crew sheet and a narrated opening sequence.
184 passing tests, and every number on screen read from the results at runtime.

What we learned

Write the evaluation rules down before you look at the results. A negative result can make a better product. And region-days on the same date aren't independent, so the confidence intervals have to resample dates.

What's next for Madrugada

A native-speaker review of the Spanish crew sheet.
A week of use with real crew leads, and their input on breaks, travel and piece rates.
Field sensors instead of airport weather.

Built during SierraHacks 2026. Weather data: Iowa Environmental Mesonet (IEM ASOS). PM2.5: U.S. EPA AirData. PurpleAir correction: Barkjohn et al. 2021 (EPA). Code written with help from Claude Code (AI coding assistant).