Saayashift

SaayaShift tells outdoor crews in Karachi the coolest time of day to work, using a real hourly heat forecast.

Comment

Inspiration

Karachi's summers are brutal, and the people who work outside in them (construction crews, delivery riders, municipal workers) usually decide when to work by habit. We kept asking a simple question: if a supervisor can only move a shift by a few hours, which hours are actually cooler? "Saaya" means shade in Urdu, and that is the idea: help crews find the shade in the day.

What it does

SaayaShift takes a real hourly forecast for Karachi, ranks the coolest complete work windows for each crew's operating hours, and shows the mean and peak apparent temperature for every option. A bilingual (English/Urdu) assistant answers planning questions using retrieved CDC/NIOSH guidance, and it refuses to answer medical or "is it safe?" questions. If the data ever goes stale, the app says so instead of guessing.

How we built it

  • Data pipeline: hourly forecasts from Open-Meteo, with every raw download hashed (SHA-256) and checked against a manifest, plus schema and range checks before anything is published.
  • Forecast model: a Ridge regression, frozen before evaluation. On our evaluation it reached an MAE of $0.342\,^\circ\mathrm{C}$ against $1.113\,^\circ\mathrm{C}$ for the "same hour 24 h ago" baseline, a reduction of $$1-\frac{0.34164}{1.11325}\approx 69\%$$
  • Window ranking: for a window $W$ of $d$ hours, we compute the mean apparent temperature $$\bar{T}(W)=\frac{1}{d}\sum_{t\in W}T_t$$ and rank by $\bar{T}$, then peak temperature, then earliest start. We deliberately use no invented "safe/unsafe" thresholds.
  • Guidance assistant: extractive retrieval over NIOSH passages, with exact citations and no generative model writing safety advice. It abstains on medical, safety-guarantee and unsupported questions.
  • Freshness rule: a forecast of age $a$ hours is treated as stale unless $0 \le a \le 3$ (or its coverage has expired), and the interface shows it.
  • Stack: FastAPI, React, pandas/scikit-learn, Neon (crew data), Render (hosting), Supabase (storage), with the pipeline on Hugging Face and GitHub Actions validating and publishing fresh data every hour.

Challenges we ran into

  • Never inventing numbers. The hardest discipline was resisting plausible-looking placeholder data. Zones and crews in the demo are clearly labelled sample data; the city forecast is real.
  • Keeping data fresh on free tiers. Free hosts sleep, disks are wiped, and scheduled jobs can be delayed. We built an atomic snapshot (rows and status always match), hash-verified syncing, a validator that refuses stale or gappy forecasts, and a backup schedule.
  • Safe AI behaviour. Making the assistant useful while still abstaining on health and safety questions took many rounds of testing.
  • Urdu support across the interface, alerts and answers without changing what the evidence supports.

What we learned

Trustworthy beats impressive: provenance hashes, honest staleness banners and refusing to answer are features, not limitations. We also learned how much of "AI deployment" is plumbing: scheduling, validation, caching and failure modes.

What's next

Real zone-level forecasts, crew feedback on whether recommended windows were practical, and SMS/WhatsApp alerts. SaayaShift is planning support, not medical advice or a safety guarantee.

Built With

Share this project:

Updates

Submission history