08/05/2026 updated
40 % partially available
Data & AI Engineer | Pipelines, RAG, Vector Search, MLOps | Python, SQL, Kubernetes
Marrakech, Morocco
Only remote
Engineering degree (Telecom Saint-Etienne, France and ENSIAS, Morocco)About me
I build the pipeline end to end: ingestion and ETL through retrieval to deployment. My hybrid search beat vector-only by 75% on recall@5 — measured, not claimed. Two years shipping production AI and data platforms in a regulated health-tech group. Native Arabic, fluent French.
Application Programming Interfaces (APIs)Artificial IntelligenceApache AirflowData IntegrationData TransformationPython (Programming Language)PostgreSQLMachine LearningMessaging Application Programming InterfaceShopifySQL DatabasesGenerative AIKubernetesMachine Learning OperationsApi Design
RAG and Knowledge Systems
Design and implementation of full retrieval stacks over document estates including contracts, tickets, wikis, SharePoint, and scanned PDFs. Hybrid dense and BM25 sparse retrieval with rank fusion and cross-encoder reranking, beating dense-only search by 75% on recall@5. Includes evaluation harnesses and CI checks to track quality over time.
Data and Ingestion Pipelines
Building robust ingestion pipelines from messy source material: scraping, OCR, layout-aware extraction, chunking, orchestration, scheduled refresh, and reconciliation. Replacement of browser-based automation with direct NTLM authentication for full corpus coverage and zero silent failures.
AI Platform and Deployment
Deployment of AI systems using Kubernetes, Helm, containerisation, and CI/CD with secrets management and resource tuning. Experience shipping offline-capable, air-gapped, auditable systems with health checks, logging, and defined failure modes in regulated enterprise environments.
Agent Tooling and MCP Servers
Development of Model Context Protocol servers connecting AI assistants to internal systems such as issue trackers, repositories, and documentation, with auth passthrough and read-only safety controls.
Python Development
Primary programming language for all AI engineering work, including pipeline development, API design, and system integration across multiple production codebases.
PostgreSQL and pgvector
Use of PostgreSQL with pgvector for vector storage and multi-tenant isolation enforced at the database level via row-level security.
Qdrant Vector Database
Production use of Qdrant for vector storage and retrieval, including HNSW index rebuilds and live embedding model migrations with zero downtime.
Demand Forecasting and ML Productionisation
Building multi-tenant Shopify forecasting apps with OAuth, verified webhooks, per-SKU forecasting, and a promotion gate that refuses to ship a model unless it beats the baseline in dollars.
OCR and Layout-Aware Document Extraction
Bundling OCR and extraction models into deployment images for air-gapped pipelines, enabling offline document ingestion with no external inference calls.
Technical Writing and Open Source
Author of How the Machines Learned to Talk, a free 34-chapter, 137,000-word course on training a GPT from scratch. Founder of Theia and maintainer of an annotated index of around 160 papers, courses, and tools.
Data Engineering and Orchestration
Batch and scheduled data pipelines in Python and SQL: Dagster and Airflow orchestration, dbt transformations, Kafka and Spark for streaming and large-scale processing, Docker for packaging. Consolidated two legacy jobs into a reusable multi-source ingestion framework feeding two knowledge bases. ETL design, incremental loading, content hashing to avoid redundant reprocessing, and reconciliation so derived stores never drift silently from their sources.
Experience
Professional experience0 - 5 years
Freelancer since2024
Languages
ArabicNative speakerEnglishFluentFrenchFluent
Project history
Founder and author of Theia. Built four complete independent AI systems including a hybrid RAG system over SEC filings with evaluation harness, a demand forecasting Shopify app optimising profit rather than accuracy, and two MCP servers built for safety. Author of How the Machines Learned to Talk, a free 34-chapter course. Runs an AI practitioner community and an annotated index of around 160 papers, courses, and tools.
AI Engineer at an enterprise health-tech SaaS group. 175 merge requests across 20 production codebases in 21 months. Built an enterprise knowledge platform from scratch in 5 weeks including offline document ingestion, hybrid search API, and UI with source citations. Replaced fragile browser-based SharePoint scraper with direct NTLM authentication. Fixed production OOM issues in ingestion workers. Performed two live embedding model migrations to Qwen3-Embedding-4B with zero downtime.