AI that does real work

• A description generator inside a live store.A PrestaShop module I wrote drafts product descriptions and SEO texts and pulls parameters out of supplier data, using OpenAI and Anthropic models. The text lands in the admin, where a human edits it before it goes live. It runs in Matrace Lema.
• An MCP server for a booking system.MCP is the protocol an assistant like Claude uses to call an application’s functions. FyzioDesk ships one with six tools: find a client, today’s appointments, free slots, book, cancel, unpaid invoices. It talks to the public API with an ordinary API key and has no route to medical records, because an assistant should only reach what a receptionist would.
• Being readable by AI search.Assistants now answer questions that used to be searches. A shop or a site needs clean structured data and a plain summary they can read: this site serves llms.txt, and one of the stores I look after has a module that generates its own.
• AI features for a long-term client.I keep building AI into the products of For Best Clients’ merchants, alongside the platform work I did there from 2019 to 2022.

I build with Claude every day: drafting code, reviewing diffs, refactoring, writing tests and documentation, and finding my way around unfamiliar codebases. It makes me faster. It doesn’t make the decisions.

• Architecture, data models and security choices are mine, and I can explain every one of them.
• Nothing ships unread. Generated code goes through the same review, tests and type checks as code I type myself.
• Client code goes into AI tools only when the client agrees, and production personal data never does.
• When a client asks for AI, I say plainly whether it will help or whether a normal feature is cheaper and more reliable.

• AI inside your product.Text that writes itself where it makes sense: descriptions, summaries, translations, extracting data out of PDFs and supplier files, sorting incoming messages. With the result always going through a person when it matters.
• An assistant that only sees what it should.The interesting part is not the model, it is the boundary around it: which data it can reach, what it is allowed to do, and what gets logged. That is ordinary engineering, and it is where most AI projects fail.
• An MCP server for your app.So your team can ask Claude or ChatGPT to do the boring parts in your own system: look something up, prepare an order, answer from your data.
• Being found by AI.Structured data, a machine-readable summary and pages that answer real questions, so assistants quote your business instead of a competitor’s.