AI that replaces real work, not just decks.

Most "AI services" sit on one end or the other: a chatbot vendor at one extreme, a consulting deck at the other. Neither solves the problem most businesses actually have, which is: we are doing the same manual work over and over, and we want it to stop

What I build sits in the middle, custom systems that read your inputs, do the language work, and hand a draft back to a human to approve. Three lanes, depending on the shape of the problem.

For repetitive multi-step processes, client intake, invoice approvals, monthly reporting, document drafting. The kind of work where the steps are predictable but the joins between them leak hours. I build the pipeline that captures, classifies, routes and actions each run, with humans approving the outputs that ship externally.

Typical compression: eight manual steps into two manual touchpoints; days of lag turned into minutes.

For tasks that need reading, thinking and writing, not just moving data from A to B. Lead research before a discovery call. Email triage. Document review against a checklist. Meeting summarisation. Tasks too unstructured for a workflow tool, too repetitive to keep doing by hand.

Each agent is single-purpose, scoped narrowly, fully observable, and always hands its output to a human to approve. The judgement stays with you; the reading and drafting comes off your plate.

For data locked in PDFs, emails, forms, web pages, anywhere structured information lives inside unstructured documents. Contract data extraction, statement parsing, web monitoring, competitor pricing pipelines. The unglamorous work of getting clean, queryable data out of messy sources.

Built with Claude's structured output, Node.js, and headless browser automation where required. The output lands in the system you already use.

The technical work behind the three lanes above. Model selection across Claude, GPT, Llama, Mistral and the open-weight families. Retrieval-augmented generation when the answer is in your documents rather than the model's training data. Private deployment for workloads that can't leave the client's infrastructure. Evaluation, observability and chaining, the practices that turn a prototype into something a business can rely on.

Not a separate purchase. The default engineering depth that comes with every build.

1. Audit, I look at what your team actually does day-to-day, in the systems you already use. Not a workshop, not a survey, a direct look at the workflow.

2. Scope, I write the spec. The exact workflow being automated, the success criteria, the stack, the timeline, the price. Fixed-fee from this point.

3. Build, I build it. Direct execution: no project managers, no junior handoffs, no brief-to-delivery translation loss. You can ask me anything technical, any time.

4. Measure, once it ships, we measure. Time saved, error rate, throughput. The numbers go into a short retrospective and an ongoing maintenance plan.

Pricing. Engagements start at £500, that's a focused fix for a single bottleneck, a few hours of work. Larger builds (multi-step agent systems, custom integrations, ongoing retainers) scale from there and are quoted per project after the scoping call. The free site audit is the cleanest way to find out which tier fits your situation.