← Work

Work · 01 of 04

Business chatbot with an evaluation lab

A chatbot for an office-furniture website, and a lab for checking every change to it.

Role
Product owner, prompt design and testing
Tools
Claude, React, Node.js, Supabase, Render
Year
2026
The Hansen office-furniture website with the chatbot buttons in the bottom-right corner, the chat bubble closed.
The closed bubble on the Hansen site

The problem

Visitors to a furniture company’s website had product questions and no quick way to ask them. The sales team wanted those conversations to arrive as leads with usable contact details, and to be able to step in themselves.

What I built

A chat bubble on the Hansen website answers visitors in their own language: I tested Romanian, English and German. It uses Claude, guided by written instructions: tone rules, a sales script, and product information from the company’s database. When a visitor leaves their details, the chat files a lead that is emailed to sales and saved. A visitor can also ask for a person, and after an escalation, an email is immediately sent to sales and on the WhatsApp account of the firm, a consultant takes over.

The chat window on the Hansen site. A visitor asks in Romanian for an office chair they need urgently, and the bot recommends one that is in stock.
Asked for a chair in a hurry, it recommends what is in stock.Captured October 2026.
The chat window after a visitor has typed an email address. The bot says the details were passed to the Hansen team, who will get in touch.
A visitor leaves contact details and the lead goes to sales.
The chat window where a visitor asks in German for a desk for their home office, and the bot answers in German.

The same chat, answering in English and German.

How I check it

A prompt change can fix one answer and quietly break another. So I built an Eval Lab. It replays saved conversations against the current instructions and shows the old and new answers side by side. When a chat goes wrong, I add it to the lab as a new case, describe the problem, and it joins a queue. Each one is turned into a fix and re-checked the same way.

A real failure from my own test: the bot answered ‘Îmi pare rău, a apărut o eroare’, and I logged it as a new case.

One fix, step by step

The bot assumed every buyer was a company and asked “which company are you from?”. The first fix made the question optional, and then the bot stopped asking at all. The second fix told it to always ask, but gently, for example “the company name, if you’re not buying as an individual”.

The Eval Lab replay screen. For one saved conversation, the original bot answer is on the left and the new answer is on the right.
Replaying the case: the old answer on the left, the new one on the right, now asking gently.
The Eval Lab dialog for reporting a problem on one bot answer, with fields for what went wrong and what it should have done.
Any answer can be reported from the lab.

The result

86 chat sessions and 14 leads since April 2026, including internal testing. 17 saved test cases to check every change.

What it can’t do yet