Stéphane Roche

Hi everyone, Stéphane here. I'm a senior TypeScript engineer (11+ years) and I add AI features to existing web and mobile products: assistants, agents with real tools, retrieval.

The product I learn the most from is Piko, a marketplace I built and run on web, iOS and Android (Nuxt, Cloudflare Workers, Supabase). Its assistant has 22 tools that read and write real data. The lesson that cost me the most: never trust the prompt with permissions. Ownership checks, quotas and cost limits all live in the code that runs each tool, so a confused or manipulated model still can't touch someone else's data.

Before that: LLM engineering at Prem, agent payments at BitGPT, and years of wallet and payments work, which is where I learned to be careful with anything that moves money.

Reading last week's mastermind recap, the "metadata layer first, MCP is only the connector" point matched what I see too. Happy to compare notes with anyone wiring agents into a product that already has users and permissions.

I also run a fixed-price two-week sprint: a working AI prototype inside your product, a demo, and an honest answer on whether it's worth going further. hello@astral-horizon-labs.com