AI Integration
Integration is everything between a model that works in a playground and a feature a customer can depend on. Calling the API is the trivial part. The work is the surface around it: streaming a response as it arrives, deciding what the product does on a timeout or a refusal, keeping conversation state somewhere sensible, and making sure a per-call cost does not become an unbounded one.
How I use it
Brand KB — grounded answers on a brand site
A knowledge base service where the model answers from a tenant’s own content rather than from training data. The integration work is grounding the response in retrieved documents and enforcing the tenant boundary in the query, so isolation does not depend on the prompt behaving.
Chat interfaces
Streaming responses token by token, holding conversation state, and handling the cases a demo never hits — a rate limit mid-response, a dropped connection, a user who navigates away with a request in flight.
Defined failure behaviour
Model calls fail, time out, and return confidently wrong answers. A feature needs a decided behaviour for each of those, because the alternative is an interface that hangs or presents a hallucination as fact.
Cost at the call site
An AI feature has a marginal cost per use, which conventional features do not. Knowing what a call costs and what happens when usage grows tenfold is part of shipping it, not something to discover on an invoice.
What I use
- Claude API
- TypeScript
- Next.js
- Model Context Protocol
Related skills
Want the detail behind any of this?
I can walk through any of it in as much depth as you want.