LLM features
Chat, drafting, summarising, extraction and classification built into an app or site, with output checked against a schema before the rest of the code uses it.
LLM features · Retrieval · Model selection
We build features powered by large language models (LLMs) into apps and sites: answers drawn from your own content, structured output that is checked before other code uses it and a server-side proxy that keeps your API keys out of the browser. We choose the model for the task, not for the latest announcement. We build our own products this way.
What we do
Chat, drafting, summarising, extraction and classification built into an app or site, with output checked against a schema before the rest of the code uses it.
Search over your own documents using embeddings (pgvector, hybrid search and reranking), so answers point back to the documents the model was given.
The model is called from your server or an edge function, never the browser, so API keys never reach it and every request can be checked, rate-limited and logged.
Mitigations for prompt injection and for wrong or invented answers, plus pre-written replies for input categories you define, so a feature declines rather than guesses when it is out of its depth.
The hosted or open-weight model picked per task on cost, quality and where the data is processed. We can explain the choice and change it later.
Where a single call is not enough: tool use, multi-step agents and workflows, scoped to one job and evaluated against it before release.
Proof
Questions
Not yet. AI is newer for us than our commerce work. We are building our own AI product and use AI agents daily in our own work. We would bring the same tested, measured method to client work.
Whichever fits the task. We compare Anthropic's Claude, OpenAI's models and open-weight models on cost, quality and where the data is processed, and we keep the choice easy to change.
We tell you before we build. Vendors differ on whether they process data in the EU and on whether they keep or train on it; for sensitive cases we can look at open-weight or self-hosted options.
No. We do build with AI-assisted tooling, but this service is about putting AI features into your product: the model, the retrieval, the proxy and the guardrails that ship with it.
Start a project
We reply with questions, a rough plan and an honest view of the effort.
Start a project