Sisu Shift
Shift tracking, pay and auto-scheduling for Finnish nurses
The problem
Nurses in Finland often get their schedule on paper or in an employer system that is hard to use on a phone. Tracking hours, checking pay and swapping shifts happen in notebooks and spreadsheets. Team leads build rosters by hand while following the Working Hours Act: 11 hours of rest between shifts, weekly hour caps and limits on consecutive days.
My wife is a nurse, and so are most of my friends. I watched them track shifts in notebooks and double-check their pay by hand, so I built the tool I wanted to give them.
Sisu Shift is a phone app with two jobs:
- For nurses: photograph a paper schedule to import it, then see hours and pay per shift.
- For team leads: collect everyone’s preferences and generate a roster that follows the law and spreads the hard shifts fairly.
Constraints
- Finnish labour rules are hard constraints. A roster that breaks the 11-hour rest rule is unusable, however good it looks.
- Pay must be exact. Premiums, tax and daylight-saving changes all affect what a nurse is owed.
- GDPR and EU hosting. The app stores employment data but no patient data.
- One developer. The architecture had to stay small enough for one person to run.
How auto-scheduling works
- The lead opens a preferences round for a period, with a deadline.
- Each member marks every day or slot as
want,canorcannot. - The lead taps Generate. The app calls a separate Python service (FastAPI).
- The service reads the team, coverage needs, absences and preferences, and builds a CP-SAT model with OR-Tools.
- Hard constraints: 11-hour rest, weekly hour caps, consecutive-day limits, coverage and absences. Soft goals: grant
wantdays, distribute shifts evenly, share weekends fairly. - The service writes the result back in one atomic database call. If solving fails, the round stays open so the lead can adjust and retry.
- The lead reviews the roster, then publishes, discards or retracts it.
Decisions
| I chose | Over | Because |
|---|---|---|
| A separate FastAPI + CP-SAT service | Solving inside the main backend | Solving is CPU-heavy. Isolation keeps the app responsive and the cost flat. |
| Labour caps configured per team | Caps hard-coded in the solver | Rules differ between collective agreements, so they come in as data with safe defaults. |
| Cloud vision (Gemini) for photo import | On-device text recognition with user-taught templates | Gemini handles loosely framed photos, so the templates and camera overlay were never needed. I dropped the on-device plan before shipping it. |
| TanStack Query with encrypted on-device storage and replay of queued changes | A custom sync engine with a conflict screen | The server stays the source of truth, and the simpler design covered the offline need. |
| Database migrations before app builds in CI | Building the app first | No user ever gets an app that queries a schema that doesn’t exist yet. |
| A pay result that stores a snapshot of the rates used | Recalculating old shifts from current rates | Old pay stays correct after rates change. |
Outcome
- Testing with real users: 20 alpha testers on Android preview builds, and 5 users through Apple’s test channel. App Store review is in progress.
- Auto-scheduler: built and tested; not yet run on a real team’s roster.
- Size: 933 commits over 9 months. About 376 app tests, 38 solver tests, 10 database test files and 5 end-to-end flows.
- Release pipeline: merging to
mainmigrates the database, then builds and submits the iOS and Android apps. - Honesty audit: I checked the project’s own docs against the code and corrected them. For example, the early plan cited HIPAA, a US law; the app actually falls under GDPR and Finnish law.
What I’d do differently
I would position it more broadly. The core of the app is not nursing: it’s fitting shift work around a life. The Today screen already tracks rest time, weekly hours and days off against the rules. That serves any shift worker trying to keep a work-life balance, such as a working parent, not only nurses. Starting that broad would have given me a larger pool of strangers to test with, instead of mostly friends.