Onboarding Assistant
The guide at the threshold — every answer sourced, every secret kept.
I build applied AI systems end to end — agents with enforced permissions, retrieval that cites its sources, and the interfaces people actually use. What you get isn't code; it's a working system and the results it produces in production.
I ship applied AI systems end to end — from agent permissioning and retrieval over private data to natural-language access to live databases. Currently an AI Engineer at SerValid, building RAG systems and LLM automations for clients; author of Custos, an open-source security library that gates every tool call an AI agent makes, and of a RAG + text-to-SQL assistant that finds the right source 91% of the time. I work where research meets reliability: experiments that must ship, and systems that keep working long after they do.
Each step up the sequence — closer to the truth, heavier the systems entrusted to me.
Designed and deployed RAG systems (FastAPI, LangChain, pgvector) that give clients source-cited answers from their own documents. Built AI automation workflows in n8n — including one that pushes new leads into a CRM in real time — and integrated OpenAI, Anthropic, Mistral and Gemini into clients' CRMs, ERPs and internal tools. Work directly with clients to turn business problems into production-ready solutions, from requirements to deployment.
Built an intranet IT intervention platform for 300+ Algérie Télécom employees, cutting average resolution time by ~40%. Delivered a real-time WebSocket dashboard with KPIs and per-department analytics, a Python/tkinter desktop widget compiled to .exe, email notifications, and Excel import/export with server-side validation. Handled intranet security: auth, upload validation, and file permissions.
Three cards from the archive — each a system shipped, each a result delivered. Turn one to read the engineering behind it.
The guide at the threshold — every answer sourced, every secret kept.
A Mistral-native onboarding assistant (FastAPI, Next.js) that answers new employees from company documents and live databases — citing its sources and showing each person only the data they're allowed to see. Hybrid retrieval (pgvector + full-text + trigram, RRF fusion) finds the right document 91% of the time; plain-English questions become read-only, department-scoped SQL.
The guardian at the gate — permission middleware for AI agents.
Open-source middleware that gatekeeps every tool call an AI agent makes. YAML policies are enforced deterministically across 7 frameworks — plugging in with zero changes to existing tool code — while a detector catches indirect prompt injection, blocks the dangerous call, and clears it from the agent's memory. Every permission decision lands in tamper-evident logs.
The contract unseen — decoded, scored, and brought into the light.
A browser extension that reads the fine print no one reads. It ingests a site's Terms of Service, runs the clauses through a lightweight on-device transformer, and surfaces a privacy-risk score before you hand over your data. No cloud calls, no trackers — a quiet oracle for the small print. Currently in progress.
Three disciplines, practised in concert — the seer reads the surface, the savant forges the foundation, the oracle awakens the machine.
Interfaces with intent — fast, accessible, precise.
Foundations that endure — observable, deterministic, scalable.
Systems that reason — models, retrieval, and orchestrated intent.