2026 – ongoing · Remote
BandS — an operations agent for a buy-and-resell activity
Entrepreneuriat personnel — Web projects run independently from 2005 to 2011, and the first sites delivered to clients.
Personal project launched in October 2026 to equip an activity of buying in clearance stores and reselling on a consumer-to-consumer platform. The goal: to know quickly and accurately whether the activity is sustainable, and to cut the time it takes. Four flows: purchases, listings, shipments and steering.
Role held: AI Product Manager — technology stack selection, setup of the AI-assisted development environment (rules written before the first line of code, one skill per flow, a dedicated operations agent), design of the business rules. The agent turns unstructured input into structured rows: a photographed till receipt becomes purchase lines and lots, photos become a listing ready to publish, a sale email becomes a transaction and a parcel preparation sheet. It publishes nothing itself — automated publishing would expose the account to a ban — and computes no management figure: margins, stock and contributions are spreadsheet formulas that it feeds.
Three gains: time given back on repetitive entry and preparation tasks, an activity that flags itself rather than on request, and financial tracking readable at a glance. Tracking relies on dashboards and monthly charts, a morning brief generated every day and a report on the first of the month. Inconsistencies are flagged, never silently fixed: a receipt whose lines do not add up to the amount paid is not written, and an ambiguous sale is not matched by guesswork. The code is written with Claude Code and covered by automated tests.
Achievements
- 4 flows equipped, from purchases to steering
Time saved by automatically structuring receipts, photos and emails
A photographed till receipt becomes purchase lines, a trip and lots; photos, a listing ready to publish; a sale email, a transaction and a parcel preparation sheet. That much data entry avoided.
- 3 financial tracking dashboards
Readable financial tracking through dedicated dashboards
Activity, account by account, and charts of past months. The figures are spreadsheet formulas, fed by the agent but never computed by it.
Human validation built into listing publication
The agent writes the listing; publishing remains a human step, since automated publishing would expose the account to a ban.
- 0 receipts recorded without a matching total
Inconsistencies flagged rather than fixed
A receipt whose lines do not add up to the amount paid is flagged, not "corrected"; a sale whose match is ambiguous is not guessed.
- 2 automated routines, daily and monthly
Proactive tracking: morning brief and monthly report with no intervention
Every morning at 6:30, a document sums up what needs action; on the first of the month, a report. The activity comes to you, with no request.
- 6 skills, all built for the project
Setup of the agentic development environment
One skill per flow — purchases, listings, shipments, steering —, a master skill, and a dedicated operations agent with its own; rules written before the first line of code.
- 32 automated tests
Test coverage of document reading
Reading of till receipts, listing drafting and sale email parsing checked by three test suites; every script that writes runs in simulation first.
Skills applied
- Agentic development environment setup
- AI Product Management
- Building skills for AI agents
- Claude Code
- Context engineering
- Data extraction from unstructured documents
- Designing specialised AI agents
- Designing steering dashboards
- E-commerce
- Functional acceptance testing and qualification
- Human-in-the-loop
- Linear
- Technology stack selection
- Third-party API integration