I build Claude-powered agents, RAG chatbots and n8n automations that are grounded in your data, measured before they go live, and keep a person in control of every decision that matters.
Animation: a replay of the AI Enquiry Agent handling a boiler-service enquiry. It classifies the message, calls four read-only tools for the customer, price, service area and free slots, drafts a reply quoting the fixed £95 price, checks it, and waits for a person to approve it before sending.
ResultsEvery project here is live. Open the demos, read the code and see the test results, misses included.
Both run on my own server, set up for a fictional plumbing company so you can use them freely: send an enquiry, ask the chatbot anything, try to trip them up.
Reads every enquiry, checks the facts, drafts the reply. A human clicks send.
Answers only from your documents, and shows you the exact passage it used.
Four core services, each backed by a working project you can try, plus the builds that usually come with them. Every project is scoped with you, tested, and handed over running in your own accounts.
Agents that take repetitive work off your team: inbox and enquiry triage, lead qualification, follow-ups, reports, data entry. Built in n8n or code, with an approval step wherever a mistake would cost you.
A chatbot for your website or team that answers only from your documents, shows its sources, admits what it doesn’t know, and hands over to a person.
Connect Claude to your CRM, database, calendar or internal APIs through the Model Context Protocol, with read-only by default and permissions you control.
Apps started in Lovable, Bolt, Replit or Cursor, finished and made production-ready: bugs fixed, logins, payments and database permissions secured, tests, monitoring and a clean deploy.
The first working version of your product: accounts, subscriptions, a dashboard and the core feature.
Data collected on a schedule from sites and APIs, cleaned and delivered to a sheet, database or API.
Invoices, forms and PDFs turned into structured data, checked and sent where you need it.
Chrome extensions, Google Sheets and Apps Script tools that save your team a few hours every week.
Your team set up with Claude Code, MCP servers and a workflow that ships reviewed, tested code.
What makes AI dependable with real customers is the engineering around the model. Every project gets all six as standard.
See them in a real buildAgents read, look things up and prepare the work. Anything that sends, books, pays or deletes waits for a click from you, unless you decide it shouldn’t.
Enquiry Agent: read-only tools, Slack approval before every send
Answers come from your data, with the source attached. If it isn’t in your documents, the AI says so and offers a person instead of improvising.
Chatbot: every claim cites its exact passage
A labelled test set built from your real cases, a written accuracy report, and the failures I found and fixed. You see the numbers before it goes live.
Enquiry Agent: 55 test cases, a held-out set, a published report
Least-privilege access, prompt-injection defences, signed webhooks, rate limits and spend caps. Secrets live in environment variables, never in code.
Both: separate database roles, each proven by script
The smallest model that passes the tests for each step, prompt caching, and the cost of every run logged. You know the price per task before you scale.
Enquiry Agent: $0.0006 to classify, $0.033 to draft
Built in your accounts, documented and handed over with a setup guide. No lock-in and no per-seat platform fees: it keeps running without me.
Both: setup guides, an “adapting this” guide, MIT licence
Every tool here is used in a project you can open and inspect. I choose established, well-supported technology, so your system is easy to run and easy to hand over.
Clear steps, a written scope and a fixed price. You always know what’s being built, what it costs and how we’ll measure it.
A free 30-minute call. We find the task that’s costing you hours and check whether AI is the right fix for it.
A clear yes or noA short written plan: what it will do, what it won’t, how we’ll measure success, the milestones and a fixed price.
Written scopeWorking software at every milestone, shown in a short demo video or call, so you can steer early instead of at the end.
Demo each milestoneWe run it against a test set of your real cases: accuracy, edge cases, injection attempts, cost and speed per run.
Evaluation reportDeployed in your accounts, with documentation, a walkthrough video and support while it settles in.
You own it allPick a package and order it on Fiverr or Upwork, or send me the details for a custom quote.
Introductory rates for early clients
AI chatbot trained on your documents. Answers customers from your own content, with sources, on your website.
$150
$400
$900
Workflow automation with AI. Connects the apps you already use so routine work runs itself.
$100
$300
$700
Fix and deploy your AI-built app. For apps started in Lovable, Bolt, Replit or Cursor that need to work in production.
$150
$450
$1,000
AI agents, MCP servers, integrations and larger builds are quoted per project after a short call, or billed at $35/hour on Upwork.
Something else on your mind? Bring it to the call.
Answers come from your own data and carry their sources. When nothing relevant is found, the AI says it doesn’t know and offers a person. Prices and other key facts are checked by code, and anything that fails a check goes to a human. Before launch I measure it on a test set of your real cases, so you see the accuracy figures before it goes live.
It’s measured per run, and usually small. I use the smallest model that passes the tests for each step and log the cost of every run: the Enquiry Agent costs about $0.03 per enquiry, so 1,000 enquiries a month is roughly $33 in AI fees (billed in USD by the AI provider). You get a cost estimate for your volume before we build, and a spending cap from day one.
In your accounts, wherever possible: your database, your email, your server or cloud. Access is least-privilege (each part can only reach what it needs), secrets stay out of the code, and I document what’s stored and for how long. Claude’s API doesn’t train on your data by default.
Yes, completely. It’s built in your accounts, the code is yours, and you get setup docs and a walkthrough so anyone competent can maintain it. No lock-in and no per-seat platform fees.
Very likely. Anything with an API or webhook can be connected: Gmail or Outlook, HubSpot, Pipedrive or Airtable, Google or Outlook Calendar, Slack or Teams, your website or help centre. If there’s no ready-made connector, I build one, often as an MCP server so Claude can use it safely.
Most packages are delivered in 2 to 10 days: a single automation in 2, a document chatbot in 3 to 10, an app fix in 2 to 5. Larger custom builds, such as an MVP or an agent connected to several systems, usually take 2 to 3 weeks, with milestones and dates in the written scope.
That’s what the discovery call is for. Tell me where your team’s hours go, such as the inbox, data entry, answering the same questions or chasing leads, and I’ll suggest the one change likely to pay off fastest.
Yes. I build with Claude Code, which is a large part of why delivery takes days rather than weeks. I design the system, review every change, and test it against real cases before you see it, so what you receive is code I understand, can explain and stand behind.
I specialise in Anthropic’s Claude (Haiku, Sonnet and Opus), which is strong at following rules, using tools and citing sources. Each step gets the cheapest model that passes its tests, and switching model is a one-line change, so you’re never stuck with one.
A fictional company I set up the demos for, so you can try them freely without touching anyone’s real data. The company is made up; the systems, the infrastructure and the test results are real and live now.
Tell me what it is. I’ll come back with whether AI is the right fix, what it would take, and what it would cost to run.
Replies within 24 hours · Free 30-minute discovery call
That’s enough for me to come back with an approach, a rough timeline and the questions that matter.