AI & ML platforms

The horizontal that cuts through every vertical

AI is a sector and also a substrate. Some clients come to us for an AI product. Others come for a marketing site and end up with AI agents in their ops. Either way, this is the capability that underpins most interesting work we’ve done in the last 24 months — and it’s the one that’s moved hardest, fastest.

What we build

• Multi-agent workflows — orchestrated pipelines where specialised agents hand off to each other, with guardrails at every seam. Built on Claude, GPT, open-weights models, or mixed. Orchestrated via LangGraph, Temporal, or plain old reliable code.
• Retrieval-augmented generation (RAG) — over your documents, knowledge bases, ticket history, product catalogue. Semantic chunking, hybrid search, re-ranking, citation, and the freshness layer most demos skip.
• Classical ML where it wins — classification, forecasting, anomaly detection. Often cheaper and more reliable than asking an LLM; we don’t use a chainsaw for a butter knife job.
• Human-in-the-loop review — queues, audit trails, disagreement logging, re-training loops. The grown-up version of “AI that’s wrong sometimes.”
• Evals, observability, cost controls — every AI build ships with an eval suite and a cost ceiling. Because “it worked last Tuesday” is not a production strategy.

What’s legitimately new (last 12 months)

• Model Context Protocol (MCP) — Anthropic’s open standard for connecting LLMs to tools. We’ve built several production MCP servers; happy to walk you through why they matter.
• Long-context models (1M+ tokens) — changes what RAG you need versus what you just cram into the prompt. We know where the line sits, per use case.
• Agentic commerce — AI agents that complete multi-step transactions on behalf of users. Early but real; a few clients are already there.
• Generative UI — UIs that adapt to the user via LLM reasoning, not a rule engine. Cached for speed, watermarked for audit.

What we won’t do

We won’t ship AI as a magic word without a testable spec. We won’t burn your OpenAI budget on a prototype that never survives a production cost-per-request check. We won’t pretend the model’s bad days don’t exist — we build for them. Reliability is what separates an AI product from an AI demo.