// journey

Journey & Experience

3 years of production AI experience, from Algeria to Paris.

// experience

Professional Journey

MapBrain

AI Engineer / Full-Stack

2025/12 → 2026/05Freelance

📍 Paris · 6 months

End-to-end design and delivery of an AI educational SaaS platform in 6 months — from multi-format ingestion to a real-time voice tutor. Full Azure architecture (AKS, Azure OpenAI), close collaboration with the product team.

PythonFlaskFastAPILangChainLangGraphAzure OpenAI+12

Zaivio

Generative AI Engineer

2024/12 → 2025/11Freelance

📍 USA (Remote) · 1 year

Design and delivery of 3 end-to-end GenAI solutions in 15 months for a US startup, fully remote. Complete autonomy on projects ranging from e-reputation monitoring to medical AI assistant.

PythonLangChainLangGraphOpenAIFastAPINext.js+5

Publicis Sapient

Data Science Consultant / AI Engineer

2024/06 → 2024/08Mission

📍 Paris · 3 months

Delivery of a generative AI tool for a CAC 40 group, achieving 90% accuracy on greenwashing detection. Collaboration with a multidisciplinary team of 10+ people.

PythonGPT-4GPT-4oPrompt EngineeringGitJira+1

Epsilon France

Data Scientist / AI Engineer

2024/04 → 2024/06Mission

📍 Paris · 3 months

Automation of RFP analysis with 50% reduction in manual work. RAG tool deployed and evaluated with rigorous quality metrics.

PythonLangChainHugging FaceAzure OpenAIRAGPrompt Engineering+1

Biopharm

Machine Learning Engineer

2023/02 → 2023/07Mission

📍 Algeria · 6 months

Delivery of a complete ML solution (model + interface) enabling precise spare parts purchase planning and reducing overstocking costs.

PythonScikit-LearnTensorFlowPandasTkinter

Jumia

Machine Learning Engineer

2022/03 → 2022/08Mission

📍 Algeria · 6 months

Seller churn prediction tool at 80% accuracy, enabling sales teams to target retention actions.

PythonScikit-LearnStreamlitPandasJavaScript

// build in public

Build in Public

Decisions, bugs, learnings and milestones in real time.

🎯

decision2026-02-15#mapbrain

Milvus vs Pinecone: why I chose self-hosted for the RAG

The choice For the MapBrain (freelance engagement) vector database, I had to choose between Pinecone (SaaS, zero config) and Milvus/Zilliz (self-hosted or serverless). Option A: Pinecone - ✅ 5-minute setup, simple API - ✅ No infra management - ❌ High cost at scale (per stored vector) - ❌ No native...

#RAG#Vector DB#Architecture#Decision

📚

learning2024-06-10

HyDE vs RAG Fusion: what I learned at Epsilon France

Context At Epsilon France, I had to optimize a RAG system for technical RFP analysis. Two advanced techniques were competing. HyDE (Hypothetical Document Embeddings) Generates a hypothetical document that would answer the question, then searches by similarity with that document. - ✅ Excellent for ...

#RAG#HyDE#RAG Fusion#Learning#Epsilon France

🔄

iteration2024-07-20

From 67% to 90%: prompt iteration for Anti-Greenwashing classification

Before — Prompt v1 (67% accuracy) General description of advertising ethical rules. The model was over-generalizing and missing subtle greenwashing formulations. Problem identified The model wasn't distinguishing nuances between: - Explicit greenwashing ("100% ecological" without proof) - Implici...

#Prompt Engineering#Iteration#Classification#Publicis Sapient

🏁

milestone2024-06-28

Milestone: 50% of manual work eliminated with RAG at Epsilon France

The result The RFP analysis RAG system went into production at Epsilon France. Result measured after 2 weeks of use: 50% reduction in manual processing time for compliance criteria. What made it possible 1. Optimized RAG pipeline (RAG Fusion > HyDE on this corpus) 2. Rigorous RAGAS evaluation befo...

#Milestone#RAG#Production#Impact#Epsilon France

// certifications

Certifications

Dataiku Core Designer

Dataiku

Dataiku ML Practitioner

Dataiku