AI Product Development

Getting the AI working is the fast part. A prototype that produces good output takes about a fortnight. The other four months go on the questions a demo never asks. What does a user see during the thirty seconds a result takes? What happens when a job fails halfway through? How do you price something whose cost to you moves every month?

Typical work looks like this. A product customers pay to use, built around a model. An internal tool that handles a volume of work you could never staff for. A demo that people liked, turned into something that can take real traffic and real money.

What you get is a live product with users on it, monitoring so you can see what it is doing, and a codebase your own team can take over.

The AI never runs inside the web request. Jobs go into a queue, a worker picks them up, the result lands in storage, and the app collects it when it is ready. Built the other way round, the site stops responding the day you get busy.

Retries are separated by cause. A model that times out gets tried again. A malformed request does not. Treat the two the same way and you can burn a month of credits overnight on work that was never going to finish.

Billing goes in at the start rather than the end. What you charge shapes how the system is built, and changing that later is expensive. On Motivia, a client platform we built and run, that is how seven generation types serve paying subscribers.

We write down what version one will not do, before we start. AI products grow without limit, because once the model is there every new feature sounds free. The list we agree on is the list we build.

We will not build on a model we have not tested against your real task first. If the output is not good enough at that stage, more engineering around it will not rescue it, and we would rather tell you in week two than month five.

Where a wrong output matters, a person approves it before it goes anywhere. You can decide to remove that later. We will not ship it that way by default.