AI consultant in London for teams that want agents, automation and AI search to ship, not sit in a deck.

Works out where a language model earns its place in your business, then builds that part and measures it. In practice: a readiness assessment of your data, workflows and risk, a written plan with a cost ceiling, then agents, automation, retrieval or AI search work shipped in sprints with an eval set behind it. The difference from an agency bench is that one senior engineer does the thinking and the building, so the plan is one that can actually be shipped.

Three starting points. An AI readiness assessment is from $2k and is credited against the build if you go ahead. A fixed-price AI feature or automation, one agent, pipeline or LLM feature with its evals, is from $4k. An agent, pipeline or RAG platform is from $10k and is quoted from a written scope. Advisory time on its own is $250 an hour; an ongoing engineering retainer is priced on the hire an AI engineer page. Every figure is a starting price, fixed in writing before work begins.

Yes, where the job is bounded: triage a queue, run an audit, draft from a brief, reconcile two systems, file the result. I build them on Claude and the Vercel AI SDK with tool use scoped to the minimum, an eval set that covers the failure modes, and hooks that gate anything destructive. Open-ended agents that are meant to run a department are not something I will sell you, because they do not work yet.

Claude by default: Sonnet for bulk work, Opus where judgement matters, Haiku for cheap classification. GPT and Gemini when the task fits them better or as a second opinion in a quality gate, which is how the humanizer on this site works. Routing goes through a gateway with failover so a provider outage is a log line, not an incident. I do not train models from scratch, and I will say so when a rule, a query or a regular expression does the job for free.

Yes, that is the most common shape of work. Search that answers rather than lists, an assistant over your own documents, extraction and classification at the edge, drafting tools inside your admin. Retrieval runs on Supabase pgvector with hybrid search and reranking, streaming goes through the Vercel AI SDK, and the feature ships with a cost ceiling and rate limits before the first user reaches it. HostList.io, Deluxe Astrology and this site all run LLM features in production on exactly that stack.

With numbers agreed before the build. Every feature gets an eval set of 50 to 100 cases that scores accuracy, citation correctness and refusal behaviour, and it runs in CI so a model upgrade cannot quietly regress it. Then the business measure: hours saved per week, tickets closed without escalation, pages published per month, citations in AI answers. The assessment names the measure; the handover report shows it.

Yes. The work is being the source that ChatGPT, Perplexity, Gemini and Google AI Overviews cite when your buyers ask. It starts with an audit of which prompts mention you today, then answer-first pages built to be quoted, entity schema that ties the brand together, and a monthly read of citation share. The AI citation checker and AI search keyword research tools on this site are free and run the first check in a minute.

London. Most clients are in the UK, Europe and the US. Calls run on UK hours with early-evening slots for US time zones, and the work itself is asynchronous: a preview link and a written note after every sprint, so nobody waits for a meeting to see progress.

Me, Gautam Khorana. I scope, build and ship the work myself, with Claude Code as the execution tool and specialists joining for legal or security review when a brief needs them. The repository sits in your GitHub organisation from the first sprint, API keys and hosting are in your accounts, and there are no licence fees on anything I build. When I leave, nothing leaves with me.

Yes, before the first call if you prefer. Your own paper or mine. Client data used during a build stays in your accounts, nothing is used to train anything, and I do not store personal data or health data on my side.

Two weeks on the problem at the top of your list before anyone writes a prompt. I look at the data you hold and how clean it is, the workflow the AI would sit in, who is accountable for its output, the privacy and residency constraints, and what it would cost to run at your volume. You leave with a written plan, a build-or-buy recommendation with the reasoning shown, and a price range whether you continue with me or not. The fee is credited against the build.

When the input is structured and the rule is known: a SQL query, a lookup table or a regular expression is faster, cheaper and auditable. When being wrong is expensive and nobody will check the output. When the volume is so high that per-token cost beats the hours saved. Roughly a third of assessments end with one of those answers, and the client keeps the fee credit for the part that does need a model.

Two weeks for the assessment. Two to four weeks for a single feature or automation with its eval set. Six to twelve weeks for an agent, pipeline or RAG platform. The part that takes time is rarely the model call; it is the eval set, the guard rails and the integration with the systems around it, and anyone promising a production agent in a week is skipping one of them.

Yes, and it is often the better shape. I run the assessment and the architecture, build the first feature with your team watching the pull requests, and leave the eval set, the Claude Code skills and the documentation so the next feature is yours to ship. Advisory time at $250 an hour covers design reviews and model decisions after that.