"The Fool that does not belong to this era,
the mysterious ruler above the grey fog,
the King of Yellow and Black who wields good luck."

Applied AI Engineer based in Paris, building production AI systems end to end — from LLM orchestration and retrieval pipelines to the interfaces people actually use.

• Author of Custos — an open-source PyPI library that enforces permissions on AI agent tool calls across 7 frameworks, with prompt-injection defense and tamper-evident audit logs.
• Built a Mistral-native RAG + text-to-SQL onboarding assistant with hybrid search (recall@5 of 0.91), ACL-aware retrieval, and a 232-question evaluation harness.
• Shipped an internal IT helpdesk platform for 300+ employees at Algerie Telecom, cutting issue resolution time by ~40%.
• CS student at CY Cergy Paris University — open to an AI engineering internship in Paris starting April 2027.
• Arabic (native) · French (fluent) · English (professional)

• Open-source security library for AI agents (Python, on PyPI) that blocks any action not explicitly allowed, checking every tool call against YAML permission rules across 7 decision types.
• Plugs into 7 popular AI frameworks (LangChain, OpenAI, Anthropic, MCP, and more) with zero changes to existing tool code — ready-made adapters, a TypeScript SDK, and a secure gateway service.
• Blocks indirect prompt-injection attacks with a detector that stops the dangerous call and clears the agent's memory of the attack.
• Proven with 600+ automated Python/TS tests and a 125-case attack simulation suite; every permission decision is stored in tamper-evident logs.

• AI chat assistant (FastAPI, Next.js, Mistral) that answers new employees' questions from company documents and live databases — every answer cites its sources, and each user only sees data they are allowed to see.
• recall@5 of 0.91 by combining three search methods — semantic, keyword, and typo-tolerant (pgvector + full-text + trigram, RRF fusion, ACL filtering).
• Employees can ask the database in plain English (NL→SQL), translated into read-only SQL scoped to each user's department.
• Backed by a 232-question evaluation harness and 87% test coverage.

"Praise the Fool."