How we built and shipped an AI SaaS in 3 months.

Three months is not a long time to ship a product. It is barely long enough to agree on a name. But constraints have a way of clarifying decisions, and a deadline that short forced us to be honest about what actually mattered: a few flows done well, and everything else deferred.

We started where most teams hesitate to — by using the half-finished thing ourselves, every day, before it was good. The roughness was the point. Each friction we hit in our own work became the next ticket, and the backlog wrote itself from real use rather than from a roadmap drawn in advance.

Build the thing you already need

The first version had no settings, no onboarding and no marketing site. It did one job: take a messy pile of documents and turn them into something a person could act on. Because we were the users, we could cut every feature that did not earn its place in a single afternoon of dogfooding.

That discipline compounded. By keeping the surface area small, we kept the codebase legible, and a legible codebase is a fast one to change. When a flow felt wrong on Monday it could be rebuilt by Wednesday, shipped to ourselves on Thursday, and judged in real work by Friday.

We leaned hard on the most capable models we could reach, then spent our engineering effort on the connective tissue around them — the parsing, the context, the guardrails — rather than reinventing the intelligence itself. The model was a component; the product was everything we built so that the model could be trusted.

Ship before you are ready

Shipping early is uncomfortable precisely because it is informative. The features we were proudest of were often ignored, and the rough edges we apologised for turned out to be the ones people loved. None of that is knowable from inside the building; it only shows up once the work meets a real workflow.

Five months on, the internal tool is a product other teams rely on. The lesson was not speed for its own sake — it was that a tight loop between building and using is the fastest way to find out what is worth building at all.