Data scientist for five years, full-stack engineer since. I shipped NLP at 1,000 req/s before it was fashionable. Now I build the products around the models — agentic pipelines, retrieval, self-hosted tools on my own time.
interface Engineer {
focus: "ml → product";
since: "2018, in production";
stack: [TypeScript, Effect, Vue, Go, Python];
rule: "a number beats an adjective";
}- notara — the notes app you actually own. A self-hosted Notion alternative: block editor, inline databases, real-time collaboration. Entire backend in Effect, one SQLite file per workspace, one container, no cloud. Architecture and licensing decisions recorded as ADRs in the repo. The largest thing I have built solo.
- permis-bateau-rag — a full RAG pipeline with a real benchmark: 60 stratified questions, 10 out-of-domain traps. The knowledge graph I spent weeks on lost to plain cosine similarity on ranking, so I cut it and wrote up why.
- herdr-automations — scheduled tasks for coding agents. A prompt, a cron line, a fresh git worktree per run. Go, prebuilt binaries, no store.
- cantine — unofficial Swile CLI. Read-only, no password, Go.
- DishNow — AI recipe and meal planning app. Nuxt 4, PocketBase, Mistral for extraction and generation, shipped to iOS and Android via Capacitor.
Full-stack engineer at FREELANCEREPUBLIK on Jemmo, an AI talent-matching SaaS: a plan-then-execute agentic pipeline, vector search over 22M profiles, and a remote MCP server exposing it to AI assistants over authenticated OAuth sessions.
thomas.legrand.sh · case studies · X · LinkedIn · Data for Good contributor