Nov 9, 2026
Seating Nearly 2,000 People Across Three Countries
How I replaced manual seating plans with one operations platform for offices in Zimbabwe, Zambia, and Costa Rica, then made it 60% faster.
The problem
ZimWorx runs offices across several high-rise floors in Zimbabwe, Zambia, and Costa Rica. Nearly 2,000 seats, each assigned to a team member who works for a specific client on that client's US timezone.
Before Atlas, all of that lived in manual seating plans. Answering "where can the new hire sit?" or "how full is the eighth floor?" meant someone counting by hand, and every move risked a double-booking.
What I built
Atlas is a single operations platform for seating, team members, and client assignments across every site.
- Interactive floor maps. Each floor is laid out on a grid and organised into zones and clusters, so a floor renders the way people actually picture it.
- Seventeen seat statuses. Occupied, vacant, faulty, equipment, and the rest, each colour-coded, because "empty" and "available" are not the same thing.
- Team member and client assignments. Every seat links to a person, and every person links to their client, with full history of who sat where and worked for whom.
- Timezone-aware schedules. Work hours are calculated from each client's timezone, so cross-site coordination doesn't depend on mental arithmetic.
- Role-based access. Admin, CTO, manager, and staff permissions on AWS Cognito, with an audit trail of every change.
The stack is a Django REST backend, a React frontend, and PostgreSQL.
Making it fast
The first version worked, but it didn't scale. Loading one floor meant fetching the floor, then its zones, then each zone's clusters, then each cluster's seats, then the people in them. That was 15 to 20 requests per page, each waiting on the last, with N+1 queries behind most of them. Under normal use we were hitting our own rate limits.
The fix was to serve data the way the UI thinks about it:
- One endpoint per floor. It returns the whole floor, zones, clusters, and seats as one tree, cached on the server for 60 seconds.
- No more N+1.
select_relatedandprefetch_relatedon every hot path. - Lean serializers. Seats carry only the fields the map draws, and clusters ship pre-computed occupancy counts.
- Client-side caching. React Query deduplicates in-flight requests and keeps recent floors warm, so moving between floors feels instant.
Results
- API calls per page load: 15–20 down to 1–2
- Page load time: 3–5 seconds down to under 2
- Database queries per load: 100+ down to about 5
- Payload size: 40% smaller
- Cache hit ratio: over 80%
Atlas runs in production across all three countries and replaced the manual seating workflow.
What I'd tell someone building the same thing
Design the API for how the screen is read, not for how the tables are stored. The database thinks in rows. The person at the screen thinks in floors.