Case studies

What it was

A platform that tracks how a brand's content surfaces in the answers of AI engines and helps marketing teams improve it. A TypeScript monorepo built with Bun: a customer-facing web app, an admin dashboard, and the services behind both.

The problem

Every answer the product shows costs a model call, and useful coverage means asking several engines the same question repeatedly over time. Done naively the bill scales with the customer's curiosity, and nobody can say which feature spent the money.

What I built

The core of the optimization engine: requests are routed to the cheapest engine that can answer them, crawled pages are cached rather than re-fetched per query, and every model call is traced through Langfuse with the feature that made it.

The rejected alternative was a single premium model for everything — simpler to reason about, and the option most teams take. It fails the moment coverage matters more than any one answer's polish, because coverage is what the customer is buying.

What it bought

Engines could be added without the spend curve bending, and cost became a per-feature number the team could argue about instead of one line on a monthly invoice. Development is ongoing.

What it was

Patient management for clinical users, with a reporting layer on top of it. The longest engagement of four years at Incora, and the one where the front end carried the most weight.

The problem

Clinical work leaves the screen. The same chart a clinician reads in the application has to appear in a document that gets printed, filed and sent — and a screenshot of a web chart is not that document.

What I built

One D3 charting layer serving both outputs: the interactive views in the application and the PDF reports are drawn from the same code and the same data, so a chart cannot say two different things.

The rejected alternative was a separate server-side report renderer. It would have shipped sooner and then drifted — two chart implementations, one of them the one nobody looks at until a clinician complains.

What it bought

New chart types arrived in the application and in the reports at once, and reporting stopped being the part of the release everybody dreaded.

What it was

A security product that watches server hardware: performance and state data arriving continuously from a fleet of machines, analysed and surfaced to operators.

The problem

Ingest never stops and analysis is bursty. One service doing both means a heavy analysis run delays the data still coming in — and in a security product, late data is the failure mode that matters.

What I built

The back-end split into services along that seam: ingest kept deliberately dumb and always available, analysis separate and scalable on its own, and a live channel pushing state to the operator view rather than waiting to be asked.

What it bought

Analysis load stopped being able to cost the product its data, and each half could be scaled and deployed without waiting for the other.