Mobile & scale
A nationwide education platform built for places with almost no connectivity
A mobile application for tracking classroom engagement and application use across a school network of more than a thousand sites — engineered to work properly in remote areas where connectivity is intermittent at best.
The challenge
The organisation needed to understand how its material was actually being used in classrooms across a very large network — engagement, usage, what was reaching students and what was not.
The difficulty was where those classrooms are. A meaningful share of the network is remote, and connectivity there ranges from intermittent to absent. A system that assumed a network connection would have collected data from the well-connected schools and been blind to precisely the ones the organisation most needed to understand.
The context
Devices might go days between usable connections. When a connection does appear it may be brief, slow and metered, so a synchronisation strategy that assumes bandwidth is as bad as one that assumes uptime.
Across a network of that size, small per-device inefficiencies aggregate into real problems.
The approach
Offline had to be the normal case, not the error case. The application was designed on the assumption that it will usually have no connection, and that having one is an opportunity rather than a requirement.
That reframing changes everything downstream: what gets stored, what gets sent, in what order, and what happens when two devices disagree.
The solution
Everything is captured and usable locally on the device. Synchronisation happens opportunistically when a connection appears, sending a compact queue of changes rather than a full state upload, and resuming rather than restarting if the connection drops mid-transfer.
On the server side, that data becomes the engagement and usage picture the organisation needs — with the important property that it covers the remote schools as well as the connected ones.
AI & automation
None so far. The current requirement is accurate measurement, and measurement is a data engineering problem.
It is, however, exactly the kind of foundation that makes later AI work viable: consistent, well-structured usage data across a large network is the prerequisite for any predictive or recommendation work, and it is the part most organisations discover they are missing after they have already commissioned the model.
Engineering
Sync and conflict resolution were the core problems. Determining what has changed since the last successful connection, ordering those changes, and resolving conflicts without ever discarding data a teacher entered.
Efficiency mattered more than usual: on a metered, slow connection the difference between sending a delta and sending state is the difference between synchronising and timing out. Devices in the field also vary enormously in age and capability, so both the storage footprint and the runtime cost had to suit the low end of that range, not the average.
Outcome
The organisation gets engagement and usage data from across its network, including the remote schools that a connectivity-dependent design would have missed entirely.
Teachers use the application normally whether or not there is a signal, which is the only reason the data exists at all.
Client, employer and product names are deliberately withheld. Happy to talk through the specifics, and the architecture, on a call.
More work
Other engagements
AI & retrieval
A knowledge assistant that answers from your own documents
Documents turned into something people can simply ask — hybrid search rather than vector similarity alone, a reranker that cuts forty candidates to six, and a numbered citation on every answer pointing at the document and page it came from.
AI agents & real-time
A voice agent that takes the call, looks it up, and answers
A caller rings, speaks normally, and gets an answer drawn from live business data — transcription running while they are still talking, a model working out what they actually want, and a query against the systems that already hold the answer.
Automation & generative AI
An unattended production pipeline: a topic goes in, a finished video comes out
A pipeline that takes a topic, researches and scripts it, generates the narration and the visuals, renders the video and files the result — with no human step in the middle. Built partly as a product and partly to learn where multi-stage generative systems fail in practice.