AI Development

Every week we meet a team that has an AI demo working on a good day and needs it working on every day. The gap between those two is the entire content of our AI practice. The demo works because a specific prompt, with a specific input shape, returns a specific output. Production is different: real users asking things the prompt didn’t anticipate, latency budgets the proof-of-concept never had to respect, error states that silently return bad content rather than loudly returning nothing, and a model provider that patches the behaviour underneath you every few months.

We build AI features that survive that transition. Our default architecture is Claude-first, because the quality at reasoning tasks and the tool-use behaviour are currently the best in the field, but we design every feature with a multi-provider abstraction so the provider choice is a config change rather than a rewrite. The Vercel AI SDK is our standard abstraction layer; when the shape of the feature needs more than it provides, we build directly on the provider APIs with an adapter we can swap.

For retrieval-augmented features, our default stack is Postgres with pgvector. Most of our clients already run Postgres, and adding pgvector avoids standing up a second database for an early-stage feature that may or may not earn its keep. When the scale genuinely warrants it, we’ll move to a specialised vector store, but we benchmark first rather than assuming. Agentic workflows get a different treatment: tool definitions designed like APIs, agent loops that can be reasoned about, and logging detailed enough to debug a failed run without rerunning it.

The part of the work that ages best is the eval infrastructure. Before we ship an AI feature, we ship the eval harness that will tell us whether the next prompt change or model update made the feature better or worse. This isn’t optional, it’s not an upsell, and it’s not something we apologise for scoping in. “It got better” is not a measurement. We build the measurement.

We’re honest with clients about where AI currently shouldn’t be the answer. Agents acting autonomously on production data are a smaller category than the demos suggest. Natural language interfaces over databases are a worse experience than a good form, more often than not. Automated customer support that’s actually good requires more scaffolding than most teams assume. If your roadmap assumes one of these and the evidence isn’t there yet, we’ll say so — and we’ll help you scope down to what will actually ship.