Python developer and data analyst, open to contract and part-time work. I build the unglamorous parts that decide whether an AI feature survives real users: the tools it calls, the rules for when it hands off to a person, and the eval set that proves it.
Recent, all public on GitHub:
- agent-sav-fr: a customer-support email agent that looks up the order and the FAQ before drafting, hands legal, GDPR and "delivered but not received" cases to a person, and sends nothing on its own. 17/17 on its eval set, every run log in the repo.
- carte-locale: an agent that turns Qloo taste data around a shop's address into a weekly plan for independent French merchants. The LLM rewrite is rejected if it drops a data point.
- rgaa-a11y-judgement: a Kaggle benchmark testing whether models can name the WCAG 2.2 criterion an HTML fragment violates.
- Guardrail and scheduling tooling for running Claude Code unattended.
Remote: Yes (async; overlaps EU hours and US-East mornings)
Willing to relocate: No
Technologies: Python (pandas, FastAPI, Playwright, pdfplumber, openpyxl), SQL (SQLite/Postgres), TypeScript/Next.js, LLM agents and tool use (Claude, OpenAI, Groq), RAG, evals, web scraping, PDF/Excel automation
Résumé/portfolio: https://yvoolab.github.io
GitHub: https://github.com/yvoolab
Email: [email protected]
Python developer and data analyst, open to contract and part-time work. I build the unglamorous parts that decide whether an AI feature survives real users: the tools it calls, the rules for when it hands off to a person, and the eval set that proves it.
Recent, all public on GitHub:
- agent-sav-fr: a customer-support email agent that looks up the order and the FAQ before drafting, hands legal, GDPR and "delivered but not received" cases to a person, and sends nothing on its own. 17/17 on its eval set, every run log in the repo.
- carte-locale: an agent that turns Qloo taste data around a shop's address into a weekly plan for independent French merchants. The LLM rewrite is rejected if it drops a data point.
- rgaa-a11y-judgement: a Kaggle benchmark testing whether models can name the WCAG 2.2 criterion an HTML fragment violates.
- Guardrail and scheduling tooling for running Claude Code unattended.
- A Notion ops workspace with 51 AI prompts and 12 SOPs, sold on Gumroad: https://yvoo.gumroad.com/l/ai-notion-workspace?utm_source=ha...
Trilingual EN/FR/ZH. Clear scope, tested code, written updates. Email two lines about the task and you'll have a fixed quote within 24 hours.