What's already running

Parents kept asking the same things — where the club runs, when it starts, whether there's a free spot. Answers were written by hand, and outside office hours people waited a day.

• A chatbot embedded directly into the client's website
• Connected to a database of 120 partner schools
• Finds the right branch from the location in the question
• Abuse protection — request limit per visitor

Invoices, contracts and ordinary correspondence all landed in one inbox. Everything was sorted by hand, retyped into banking software and watched so nothing slipped through.

• Document type recognition — invoice, contract, other
• Supplier identification and matching to the right company
• Automatic generation of the payment order
• Handled mail gets labelled straight in the inbox
• Telegram alerts for items that need a human look

Most incoming questions repeat, but the answer has to sound like a specific person — not a template. Writing them took hours every week.

• Connection to the mailbox over IMAP
• Learns the tone from previously sent mail (RAG)
• The generated draft waits in a folder, ready to send
• Handled e-mails get labelled automatically

Nothing ever goes out in the client's name without them knowing. The system saves the writing, but the final word is always human.

Payments arrived across four accounts at three different banks and were logged by hand into a shared spreadsheet with dozens of tabs. The catch: the payments carry no reference numbers, and the payer is often named differently from the participant — a parent pays for a child, a grandparent for a grandchild. Matching them meant reading the note on every single payment.

• Reading and parsing statements from three banks
• Extracting the payer, amount, date and payment note
• Matching against the participant roster — by name, by note, or both combined
• Finding the right tab out of dozens and writing to the right row
• A confidence score on every match; uncertain cases go to review
• Detection of duplicates and payments that don't belong in the records

The client needed an assistant for routine work but didn't want their mail, documents and personal data travelling to a third party's servers. That ruled out the usual cloud tools.

• Orchestration in n8n and a Python core — both on the client's machine
• Controlled by an ordinary message, no new interface to learn
• Reads and sorts mail, prepares draft replies
• Works with files only inside defined folders (safety whitelist)
• Watches for new listings matching set criteria
• An adaptive layer — the agent gradually builds its own knowledge base

The core and individual modules are finished and run as auto-starting services. Deployment and tuning at the client are currently under way.

It isn't a service I sell. It is, however, the clearest demonstration of how I build a system that has to run non-stop, survive failures and stay measurable — which is exactly what business automation needs.

• Real-time processing of exchange data
• History database and a task queue
• Dashboard with status and metrics
• Runs in test mode — no real money involved