Hospital platform: notifications, a permissioned Drive and locked rosters
Operations software for a hospital in the Netherlands: rosters and leave, tasks and tickets, patient information and a shared Drive for documents.
- Role
- Backend engineer
- Timeline
- Dec 2025 to Apr 2026
- Status
- Live at a hospital in the Netherlands
- Stack
- NestJS, GraphQL, TypeORM, Postgres, S3
- Problem
- Notifications weren't reaching staff, documents needed Google Drive-style sharing that stayed safe for patient data, and confirmed rosters could still be edited.
- What I did
- As one of two backend engineers, built real-time notifications over GraphQL subscriptions, rebuilt Drive with inherited permissions, external sharing and an activity log, and added roster period locking with end-to-end tests.
- Result
- Live and in use at the hospital.
The system
A hospital in the Netherlands runs its day-to-day operations on this platform: staff rosters and leave, tasks and tickets, patient information and a shared Drive for documents. Patient data and European privacy law set the bar. Every file and record has to be visible to exactly the right people, and access has to be traceable.
I joined as one of two backend engineers on a team of seven, with two frontend engineers, a mobile developer, a QA engineer and a lead. The backend is NestJS with GraphQL, TypeORM on Postgres, and S3 for files.
Backend engineer, 2025 to 2026. I owned the real-time notification system, the rebuild of Drive (permissions, sharing, activity tracking and patient files) and roster period locking, and built the Patient Information, ticket and DPIA modules. I documented each module's API and worked alongside the frontend, mobile and QA engineers through integration. The code belongs to the client, so this write-up draws on my work log rather than the repository.
- in hospital use
- Live
- systems owned
- 3
- person team
- 7
The parts I worked on. Hover or tap a node.
Notifications that reach the right person
A manager approves a nurse's leave.
My first big task was finding out why notifications never arrived in the roster and Drive parts of the app. There was no real-time path at all, so I built one on GraphQL subscriptions: a single shared publisher, a notificationAdded subscription that only delivers events addressed to the signed-in user, and support for both WebSocket protocols so older and newer clients could connect.
Saving a notification publishes it once. From there I wired in the events staff actually care about: leave requests and approvals, shift updates and reassignments, and comments on tickets, tasks and documents. A round of targeting fixes with the frontend team made sure each event reached exactly the people it concerned.
A Drive with permissions you can reason about
A nurse opens a rota in her ward's folder.
The document system had to work like Google Drive while holding hospital documents, so I rebuilt it from the data model up. Department drives and a public drive are kept apart, and rights (view, upload, edit, delete) are granted on folders and inherited by everything inside them, so access is decided in one place rather than file by file.
Operations that could quietly change access are checked. Moving a folder is validated so it can't end up inside itself or somewhere the user can't reach, and moving a file into the public drive strips its old grants except edit, so a move never widens access by accident. Around that sit the everyday features: starred items, trash, copying files and folders, file counts across nested folders, and downloading a whole folder.
Sharing outside the hospital, and taking it back
Some documents have to leave the building. Files and folders can be shared through a public link, optionally protected by credentials, or with people outside the system. I replaced the old access-type flag with share scopes that describe who a link is for, and removed the obsolete columns with a migration, so creating, checking and revoking a share follow one rule.
Revoking a share cuts off external access immediately. Every share, copy and rename is written to an activity log with its context, along with view and download counts, so there is a trail of who did what with which file. Unit tests cover the activity service and its listener.
Patients got their own folders in Drive, scoped so files linked to a patient are only visible to the people allowed to see them.
Rosters that stay put once confirmed
A confirmed roster shouldn't change by accident. I added period locking: a roster admin locks a confirmed period, and the API then rejects any shift being created, edited or deleted inside it, so the lock holds whichever client sends the request. Locked periods load read-only in the app.
I also wrote end-to-end tests for the roster's core flows: moving between periods, open shifts, bulk upload that saves everything or nothing, repeating shifts edited one occurrence at a time or as a whole series, locking, and combined area, room and department filters, checked so nobody sees shifts outside their permissions.
Privacy as a feature
Healthcare in Europe runs under the GDPR, so the platform needed more than access control. I built the DPIA module, where the hospital records data protection impact assessments with their risks and the measures that mitigate them, seeded with an initial set through a migration.
I also built the Patient Information module and the ticket feature for requesting resources (from data model and ER diagram to API and documentation), added pre-parental leave, and upgraded vulnerable dependencies across the project.
Decisions worth defending
One publisher, filtered per user
Every notification goes through the same channel and the subscription filters by recipient, so a new kind of event is one publish call, not a new stream.
In-memory pub/sub on one server
The API ran as a single server, so an in-process publisher was enough. Moving to a Redis-backed one is the step for when a second server is added.
Share scopes over an access-type flag
A scope says who a link is for, which turns creating, checking and revoking shares into one rule instead of special cases.
Locks enforced in the API
A locked period rejects writes on the server, not just in the interface, so no client can change a confirmed roster.
In their words
When Emjay joined, we had a project with several key features still outstanding, including Drive and patient-facing modules. He took full ownership of those builds and immediately raised our speed of delivery. Day to day, Emjay is highly independent: he owns features end to end, delivers on time, and asks only the questions that matter. I’d recommend him without hesitation for any full-stack development work.
ThankGod Ajayi