Case study · AI engineering
LeBread Ops
A Telegram assistant for staff admin at LeBread, a bakery café in Bern.
- Client: LeBread, Bern
- Role: everything, from requirements with management to rollout
- Status: in production
- Python · SQLite · APScheduler · Claude API (tool use) · Telegram Bot API · Google APIs
The problem
Overtime lived on a paper sheet. Every month the manager went through it by hand and checked the hours. That took about 4 hours.
What it does
Staff use it in Telegram. They can:
- log their working hours
- swap shifts
- report sick
- request days off
Admins approve requests in the same chat.
Monthly statements are saved to Google Drive as PDFs.
Shifts and hours sync to Google Calendar and Sheets.
Reminders go out on schedule, and the database is backed up every night.
The design rule
The AI chooses. The code does the work.
The language model reads the chat and only picks which tool to call. Pay and hours are calculated in plain Python and SQL. Any change to the data is a proposal, and a person confirms it before it's saved. The monthly reports run with zero AI calls.
How a message moves through the system
The LLM only picks the tool. Python does the maths, and a person confirms every change to the data.
Numbers
- saved on overtime checks (measured)
- 4 h/month
- AI cost per request vs v1*
- −90%
- automated tests
- 1,700+
- test coverage
- 99%
AI cost per request
Relative cost, v1 = 100.
* Cost per request: early numbers, being measured.
Code
The code belongs to a client project and stays private. I'm happy to walk through it in an interview.