Preamble
A working notebook kept by a developer who treats artificial intelligence as an engineering material - not a promise.
I ship AI into production web apps - for teams that need to automate, scale, and deliver smarter interfaces. End to end, from embedding to pixel.
user - context
text-embedding-3
pgvector - hybrid + rerank
Claude - JSON tools
Zod - evals + LLM judge
Catalogue of instruments
AI reconciliation - freelance client work
Reconcile volatile external sources (websites, maps, PDFs) with an internal database of thousands of physical spaces, without the AI ever hallucinating deletions. The deterministic engine decides; the LLM reviews and can only downgrade.
Stack - TanStack - Drizzle + Postgres - OpenAI - Zod - Turborepo
Legaltech - production
Legal-document checker wired into Legifrance and Judilibre. Structured extraction, citation verification, alerts on outdated case law. Built for firms that can't afford to be wrong.
Stack - Next.js - FastAPI - pgvector - Anthropic - RAG
Ops - B2B SaaS (client engagement)
Project management whose reporting writes itself: weekly digest, drift detection, client brief. The team stays in their tool, the board gets a readable brief every Monday.
Stack - Next.js - Node - Postgres - OpenAI - Stripe
Mobile + Admin - client work
An exam-prep platform for maritime certifications: an iOS/Android app and a web admin, designed and shipped end to end. Timed quizzes, mock exams, rich pedagogical content, and a drag-and-drop admin workflow to orchestrate the whole thing.
Stack - Expo / React Native - Next.js - Supabase - dnd-kit - Zod
Methodology and references available on request - contact@pauldosser.fr.
AI integration and product engineering engagements for scale-ups and studios: audit, prototype, production rollout, team support.
Leading a product and engineering team: scoping, architecture, delivery, technical coaching. Focus on AI integration and production reliability.
Public Procurement Data Space (PPDS) project for the European Commission: SPARQL queries, Virtuoso knowledge graphs, front-end and back-end work supporting public data accessibility and transparency.
Teaching modern web architecture, supervising student projects, engineering practice.
Certification platform: product, API, client-facing interfaces. First scaling experience, first real production incidents.
I sit inside your product for two to three days. I leave with a testable hypothesis, a named risk, a demo that's doable in two weeks.
Short sprints, weekly demos, automated evaluations from the first commit. We measure before celebrating - and we keep the numbers.
Deployment, monitoring, handover to your team - or ongoing support. The goal: the system holds on a Monday morning without me.
“A good AI integrator spends 70% of their time on what surrounds the model - data, guardrails, interface, evaluation. The rest is a matter of taste.”
What people ask first
The five questions that come up most often before a first call. If yours is missing, write directly - I always reply.
A letter
If you have a product that could reason a little better, a process a model could lift off your team, or an AI hypothesis waiting for proof - write. I answer within 48 hours with an honest, free, and often more useful-than-expected opinion.
- Paul.