An operations platform for bike rentals and guided tours

BikeNorth · Denmark · 2025–now

BikeNorth is an international bike rental and guided-tour business. It needed one system to run the whole company, and one that could follow it into new countries.

In numbers

• Time: about 10 seconds to confirm an agency booking instead of 5 minutes to type it, which saves around 190 staff hours a season. Bike handovers went from 4 minutes to about 1.
• Money: about €6,500 of staff time saved per season, for a server that cost €1,800 once, and no fee per email.
• Errors: the model reads 96% of fields right and marks the rest for a person; about 3% of hand-typed bookings had a mistake. Double bookings went from 14 in a season to 0.
• Adoption: 12 staff use it every day, and a new staff member learns it in about 45 minutes.
• Scale: 8,000 bookings a season, 2,400 of them by email from 35 travel agencies, and 350 bikes in 2 countries.

What I built

• Online booking for customers.
• QR-code handovers: staff hand bikes over by scanning a code.
• Automatic agency bookings: bookings from travel agencies arrive by email and are recognised by an AI language model that runs locally, not through a cloud AI service.
• People stay in control: staff confirm or decline each booking the AI suggests.
• Run by the people who use it: designed so the staff who run the business can manage the system themselves.
• Ready for new markets: designed from the start to expand into new countries.

The emails nobody wanted to type

Travel agencies send bookings the way they like: a table in one email, three lines of chat in the next, half in German, the date written four different ways. Before the platform, staff copied each one into the booking system by hand, about 5 minutes per email in high season.

The first version read the emails with Google’s Gemini API. It worked, but every email left the company, and every request cost money.

From a cloud API to a model that never leaves the building.

So I moved it in-house. I fine-tuned a small open language model on 1,200 of BikeNorth’s past agency emails and the bookings they became. It runs on the company’s own server, reads an email in under a second and fills in the agency, date, tour, guests, bikes and guide language.

Why staff still press Confirm

The model gets 96% of fields right. The rest is why it only suggests. Each suggestion shows how sure the model is about every field, and anything uncertain is marked for a person to check. Staff confirm a booking in about 10 seconds instead of typing it for 5 minutes, and nothing is booked without them.

My role

I owned the product end to end, from the domain model and workflows to the architecture and a working system that covers every part of the business.