Reliable AI features inside your SaaS.

An AI product is software whose main workflow depends on a model. An AI feature is one capability — search, an assistant, classification — added to a product that already has accounts, data, and a job to do. Most of our work is the second. Both are quoted as a fixed scope.

We pick from the providers that fit the task, usually OpenAI or Anthropic, called from your backend. The choice is written in the scope and is based on latency, cost, and what data you are willing to send. We do not train a foundation model, and we can change provider later if the agreement covers that work.

Before launch we agree example inputs and what an acceptable answer looks like: relevance, citations, or a correct category. We review failures on that set. That is a check, not a guarantee the model will be right for every future input.

Provider keys stay on the server. Each request is tied to an authenticated user and only the data that user may see. We log calls and set a spend limit or alert. Ongoing inference cost is yours to pay the provider; we document the expected pattern in the handoff, and we do not promise a flat AI bill.

The controls in a typical scope are: the provider key stays on the server, each call is tied to an authenticated user, the model only receives data that user may see, requests are rate-limited and logged, and an action that changes records or money needs a person to confirm it. We also define a fallback when the provider errors. These controls reduce bad outputs and surprise bills. They do not make answers accurate, complete, or safe for every input.

Yes. The quote names the features, the data sources, the evaluation examples, and what “good enough to launch” means for that set. Ongoing inference charges from the model provider are yours. A later feature is a new quote.