Case study · E-commerce · AI review analysis
Onebird: 5,869 competitor reviews turned into a spec sheet a factory builds from
Onebird designs its own products, has them made in Asia and sells them on bol.com, the largest online marketplace in the Netherlands and Belgium. We built the application that reads competitors' reviews, scores what buyers complain about and praise, and writes the sourcing document Onebird takes to the factory.
- Client
- Onebird B.V., Venlo, the Netherlands
- Market
- bol.com, the Netherlands and Belgium
- Engagement
- September to November 2026
- What we built
- Review analysis, scoring, a sourcing document builder and an Excel export
- Status
- Live on Onebird's own server since 18 September 2026
- 5,869reviews read in the first full production run, across 27 products
- 6steps from a category link to an Excel sourcing document
- 0AI calls inside a score, so the same reviews give the same score every time
- 1,800+automated tests on the backend
The problem
Onebird's position is the lowest price with the best specifications. The answer to “what should we build?” was in competitors' reviews: hundreds per product, in Dutch, read one by one by the CEO.
- It did not scale. Onebird launches eight products a quarter.
- It was not repeatable. Scoring reviews with an AI chat assistant gave a different answer on every run.
- Stars hide half the signal. A one-star review often praises something. “Good suction, but the battery is broken” is one complaint and one compliment.
“I could not really trust the data every time.”
What we built
The model labels. Arithmetic scores.
The AI reads one review at a time and returns only which topics it is about, each as a complaint or as praise. Those labels are stored. The impact score is a plain calculation over the stored labels, with no AI involved, so it can be tested and it never drifts.
Complaints weigh more the fewer stars the review gave.
A complaint in a one-star review counts five times as much as one in a five-star review. The weighting can be changed in Settings, and every saved weighting is kept as a version.
Targets for the factory.
For each specification, the app works out an ideal target and two alternatives from what the best sellers offer, with a priority and a remark per row.
Export to Onebird's own template.
One button fills the company's Excel template. Every export is kept as a numbered version and cannot be edited afterwards, so what went to the factory can always be opened as it was.
From one review to one line in the spec sheet
How the AI is kept in check
- It only labels. It never produces a score or a number.
- Its words are marked. Every AI-written text carries the label “Written by the AI”.
- A guard removes AI text that states a figure. Every number on screen is counted by the app.
- It sees only the review text and the stars. The reviewer's city is removed before anything is saved.
- A person chooses. The brand manager picks the product group to source in and checks each target.
Results
| Before | Now |
|---|---|
| Hundreds of reviews per product read by hand by the CEO | 5,869 reviews across 27 products read in one run |
| A different answer each time from an AI chat assistant | The same reviews always give the same score |
| Star ratings as the only measure | Complaints and praise counted separately, per topic |
| A sourcing document written by hand | An Excel file in Onebird's own template from one button |
First build to the final design in seven weeks, from 11 September to 31 October 2026.
Under the hood: stack and delivery
Stack: Python, FastAPI, SQLAlchemy and PostgreSQL. Next.js for the interface, with live updates. Docker on the client's own server in Frankfurt, with the full test suite running before every deploy.
Model choice: ten AI models were tested on 50 real Dutch reviews, and one was chosen for accuracy, consistency and cost. It is an open-weight model.
Export check: each export is compared with Onebird's Excel template cell by cell: values, number formats, fonts, fills and merges.
Still reading reviews by hand to decide what to make next?
Tell us where your product decisions come from today on a 30-minute call. You leave with a plan and a cost, in writing.
Book a discovery call