Care-home management platform

The client was a UK startup building software for child care homes. I was the engineer on the pilot, and I designed, built and shipped the first version. The pilot ran for a month, and the product then ran for about a year as the live system.

As the product moved to production I stayed the technical owner, setting the architecture, the cloud setup and the security model. It is the broadest project I have worked on: product conversations, architecture, infrastructure, security and delivery, all in one.

The notes app lets home managers record the notes they take on each child every day, by typing or by speaking. Voice recordings are saved and converted to text with Azure Speech, and AI summarises the notes and produces insights on a child’s behaviour across a week or a month.

Notes can be personal, home-level or child-level, and roles for managers and admins decide who can see and change what. Edit locking stops two people overwriting the same note, reports can be scheduled, and the real-time pieces, locking and online presence, run over WebSockets.

Care homes receive referrals for children by email. I built the pipeline that brings those emails into the app. A dedicated email service syncs Outlook mailboxes through the Microsoft Graph API, which needed a demanding authentication setup, then preprocesses and tags each message and passes it to the referrals service over pub/sub.

There, an LLM on Azure OpenAI extracts the key information. Langfuse holds the prompts and provides tracing and monitoring for every call, and the portal lays it out so managers can decide quickly. From the portal they can comment on a referral, mention colleagues and act on the email without leaving the app.

Referrals were the more complex of the two products. The client was an early-stage startup, so requirements changed frequently, and I adapted the design as what managers needed became clearer.

Access is organised as organization, then homes, then users, with multi-tenancy and role-based access control on top of Auth0. I designed the auth service end to end, and every backend service relies on it instead of reimplementing access rules.

Users are cached in Redis, so repeat requests skip a round trip to the identity provider.

The backend is a set of FastAPI services with a controller–manager structure, standardized API schemas and consistent exception handling. Redis pub/sub connects services, and PostgreSQL holds the data, where schema design and a few complex database problems were a large part of the work.

Everything is containerized with Docker and deployed by GitHub Actions to Dev and Prod on Azure VMs, behind Application Gateway, a load balancer and a web application firewall. I also maintained the VMs, handled production load and ran the production deployments.