Work

Client

A mobile car valeting business, run by an owner who spends the day out on jobs.

Problem

When your hands are busy with a customer's car, you can't answer the phone. A missed call from a new customer often goes to the next business they find. Calling everyone back in the evening takes time, and the job may already have gone elsewhere.

What was built

1. A phone assistant that answers when the owner can't. It says up front that it is an AI assistant, and recognises returning customers by their phone number.
2. It quotes only from the business's own price list and service descriptions. It checks the postcode against the area covered, and says when a larger or very dirty vehicle may cost more.
3. It checks the owner's calendar for free times, allowing for how long each service takes and a gap between jobs. It offers up to three times and books only after the caller confirms the service, day, time, address and price.
4. Returning customers can move or cancel a booking. The assistant confirms who they are first and records the change for the owner.
5. Complaints, unusual jobs and anything it can't answer from the business's own information become a message for the owner to call back. Every call is logged in a spreadsheet with a summary and the full transcript.
6. It runs on a shared engine. Each business is one settings file, and every change is tested on test calls before it reaches a business.

Tools

A voice AI platform (Vapi) with GPT-4.1, a natural-sounding voice (Cartesia), speech recognition (AssemblyAI), n8n workflows, Google Calendar and Sheets, and Python to build and deploy each business's setup.

Result

The full booking flow works on test calls, from the first hello to a confirmed job in the calendar. Each fix found in testing became a new engine release. The same setup suits any business that books jobs by phone: trades, cleaners, groomers and salons.

Tested on calls to a test calendar, October 2026. Not yet answering real customers.

Client

A consultancy company: the enquiry form on its website.

Problem

A web form usually ends in a bare email: a name, a company and a few lines about the problem. Before a useful first call, someone has to look the business up, work out what it does, think through the problem, and copy the details into a list so nothing is forgotten. That research is repeated for every enquiry, and enquiries that arrive on a busy day can sit unanswered.

What was built

1. The form sends each enquiry to an automation running in the cloud. It checks the details, filters out spam and saves the submission straight away, so no enquiry is ever lost.
2. It then researches the business: the official company register for its status, age and industry, and a web search for what it does and what people say about it.
3. An AI model reads the enquiry alongside that research and writes a short briefing: who the business is, what is really behind the problem, which parts suit automation, a sensible first project and the questions to ask on the first call. Contact details are left out of what the AI and the search see.
4. The briefing is saved as a document in a shared Google Drive folder, and a new row is added to a leads spreadsheet that works as a simple database, with a status and a follow-up date for each enquiry.
5. An email arrives with the summary and links. If any step fails, the email says which one, includes everything the visitor sent, and the rest of the steps still run. A weekly check tests every connection and only sends an email if something needs attention.

Tools

A web form, Python on a serverless cloud platform (Modal), the Companies House API, a web search API (Tavily), Claude as the AI model, Google Drive and Sheets, and a transactional email service.

Result

Each enquiry turns into a researched briefing, a spreadsheet row and an email in a couple of minutes, with no copying or searching by hand. The same pattern suits any business that takes enquiries, quote requests or applications through a website.

Tested end to end on the live site, 25 October 2026.

Client

A property business, plus a consultancy company's mailbox.

Problem

Years of unsorted mail in two business mailboxes: agent updates, supplier invoices, finance, legal and property-portal alerts, all in one place. Important admin, such as licensing letters and remortgage paperwork, was easy to miss.

What was built

1. A sorting system that reads each message's sender and subject and files it into property and topic folders, using rules I approved. It copies, checks, then removes the original. It never deletes, and it logs every move.
2. A date finder that spots dated items such as inspections and deadlines and adds them to Google Calendar, with a note pointing back to the email.
3. A weekly routine every Sunday evening that sorts new mail, updates the calendar and writes a one-page brief of the week ahead.

Tools

Python, IMAP email, Google Calendar, Claude as the AI agent, and a scheduled weekly task.

Result

The first run filed 8,373 emails into 19 folders across the two mailboxes, with nothing deleted and every move logged. It now runs on its own every Sunday.

Figures from the system's own move logs, 19 October 2026.