Hamza Ouhssaine
Contact

Hamza Ouhssaine

GenAI engineer and senior data scientist in Paris. I build LLM agents and RAG that compute rather than guess, and cite their sources.

Experience

GenAI for a bank today, after computer vision for a steelmaker and big data for the airline industry. One folder per chapter, the latest on top.

BPCE & Natixis

GenAI for finance 2026 – present

The two leaning glass towers of Tours Duo in Paris, the taller one carrying the Groupe BPCE sign, above railway tracks.
Tours Duo, Paris 13e, headquarters of Natixis and Groupe BPCE.

BPCE & Natixis Data Scientist, GenAI

– present, Paris

  • Steering ALM Metrics. Platform automating liquidity (LCR) monitoring and reporting across Groupe BPCE, built and deployed with Treasury, ALM and IT teams to analyze large-scale balance-sheet data and explain each change. An LLM-powered AI agent answers analysts’ questions in plain language and drills into the data.
  • Insight Generator. Financial data analysis and business intelligence platform for Natixis. Users import their files, explore them with pivot tables and charts, and export reports. A generative AI assistant (LLM, RAG) helps teams extract insights and improve business performance. How Insight Generator works
  • AAR (Accounting Analytical Review). App streamlining quarterly statutory and consolidated accounting reviews for about 200 Natixis entities. It drafts analyses of significant account movements and Audit Committee report sections for reviewers to edit and validate. Used by Permanent Controls and Consolidation teams to cut manual effort and standardize a process behind management decisions and audits.
  • APAC T&E Policy Chatbot. Generative AI chatbot (LLM, RAG) for the APAC travel and entertainment (T&E) policy. It automates answers to employees’ questions, grounded in official guidelines with links to sources. Administrators add documents, set custom rules and review feedback from a dedicated dashboard.

Tools Generative AI, LLMs, RAG, LangGraph, LangChain, ChromaDB, Python, Machine Learning, Microsoft Graph API, AngularJS, JavaScript, Docker, Jenkins, OpenShift

ArcelorMittal

Computer vision on the steel line 2022 – 2025

Wrapped steel coils standing on their sides in a long industrial hall, beside heavy press tooling.
Steel coils in a mill warehouse: the kind of product whose IDs the vision pipeline reads.

ArcelorMittal Data Scientist

– , Marseille

  • Product ID Recognition. Computer-vision pipelines reading steel product IDs at a 0.098% character error rate, reducing manual checks and improving traceability.
  • Predictive models for steel production. Cooling-control models improving production reliability, and computer-vision defect detection on casting rolls and products, integrated with cross-functional teams to cut manual inspection.
  • MLOps on Google Cloud. Set up the Vertex AI side end to end (Workbench, custom training, Pipelines, Model Registry, online and batch prediction, Gemini-based RAG) for product-ID OCR, defect detection and an LLM/RAG assistant, in a hybrid setup: real-time inference in the plant, training and batch jobs on GCP.

Tools Python, PyTorch, Deep Learning, Computer Vision, OpenCV, Image Processing, LLMs, RAG, Vertex AI, Elasticsearch, Kibana, Dataiku, Decision Sciences

Amadeus

Big data for airlines 2021 – 2022

A split-flap airport departures board listing flights to Singapore, Auckland, Los Angeles, Bangkok, Hong Kong and Kuala Lumpur.
A departures board: the seats airline revenue management prices.

Amadeus Big Data Engineer

– , Nice

  • Built and maintained production ETL pipelines over large-scale airline datasets for Revenue Management, with automated testing and deployment.
  • Delivered projects for Southwest, Polish Airlines, Singapore Airlines and Kuwait Airways; served as team leader and led hiring for the team.
Meelo · CRIStAL

Research and first models 2019 – 2020

Meelo Data Scientist Intern

– , Lille

  • Optical character recognition for specific administrative documents, with state-of-the-art document OCR built on LSTM, CNN and Transformer networks.
  • An API that recognises companies’ tax reports and computes their financial score.
  • Fraud and non-payment detection for consumer credit, with machine learning on open data from INSEE.
  • Corrective maintenance of the prediction models in production.

Tools PyTorch, Machine Learning, OCR, LSTM, CNN, Transformers

CRIStAL Lab Deep Learning Researcher

– , Lille

  • Deep learning for steganalysis: detecting hidden messages in images with a deep residual network, ResNet-12.
  • Channel pruning of the ResNet-12 with CALPA-Net.

Tools PyTorch, Image Processing, ResNet, CALPA-Net

Centrale Lille

Mathematics and engineering 2014 – 2020

The main building of Centrale Lille on the Cité scientifique campus: a glass front and a brick gable lettered “ECOLE CENTRALE DE LILLE”.
Centrale Lille, Villeneuve-d’Ascq.

Education

  1. 2016 – 2020

    Engineer’s Degree, Data Science & AI, Centrale Lille

    Master’s-level engineering degree from one of France’s leading Grandes Écoles, part of the Écoles Centrale group: two years of generalist, multidisciplinary engineering training, then a specialization in data science and AI taught by researchers and industry professionals.

    Coursework and skills
    • Machine learning
    • Big data
    • Statistical modeling
    • Data-driven decision-making
    • Applied mathematics
    • Optimization
    • Relational databases (SQL)
    • Project management
    • Innovation
    • Generalist core in physics, engineering sciences, computer science and economics
    • Data science integration project

    Skills Data Science, Big Data, Deep Learning, Reinforcement Learning, LLMs, RAG, Dataiku

  2. 2019 – 2020

    Master 2 (M2), Applied Mathematics, University of Lille

    Master’s degree in applied mathematics completed in one year in parallel with the engineering degree, through a partnership between the University of Lille and Centrale Lille.

    Coursework
    • Stochastic processes: modeling random phenomena that evolve over time
    • Concentration of measure: high-dimensional probability, the theory behind generalization guarantees in machine learning

  3. 2017

    Bachelor’s Degree, Mathematics, University of Lille

    Applied mathematics degree taken in parallel with the first year at Centrale Lille, through a partnership between the University of Lille and Centrale Lille.

    Coursework
    • Topology
    • Probability
    • Differential equations
    • Differential calculus

  4. 2014 – 2016

    Classes Préparatoires MPSI / MP*, Ibn Ghazi

    Two-year intensive mathematics and physics program with weekly oral exams, preparing for the competitive entrance exams to France’s top engineering schools. Admitted to Centrale Lille via the Centrale-Supélec exam.

    Coursework
    • MPSI: analysis, algebra, finite probability, inner product spaces, linear differential equations
    • MP* (advanced “star” class): matrix reduction and the spectral theorem, normed spaces, power series, parameter integrals, differential calculus, discrete random variables and the law of large numbers
    • Also physics, chemistry, engineering sciences and computer science (Python)

Certifications

  • Build and Modernize Applications With Generative AI Google Cloud Skills Boost, Feb 2026
  • Advanced Machine Learning on Google Cloud Specialization Coursera, May 2021
  • Data Engineering, Big Data, and Machine Learning on GCP Specialization Coursera, Mar 2021
  • Certified SAFe® 5 Practitioner Scaled Agile, Inc., Dec 2021 – Dec 2022
  • MOOC “Gestion de Projet” (project management) Centrale Lille
  • TOEIC TOEIC® Program

Languages

French
Fluent
English
Fluent
Arabic
Fluent
Tamazight
Native

How I build

One question, followed through the kind of agent I build. The model plans and explains, code computes, and nothing is answered without evidence.

One question through the kind of agent I build
Question from a finance analyst Plan the LLM picks tools Run SQL exact figures, DuckDB Search documents BM25 + vectors + rerank Check enough evidence? Cited answer figures and sources Question from a finance analyst Plan the LLM picks tools Run SQL exact figures, DuckDB Search documents BM25 + vectors + rerank Check enough evidence? Cited answer figures and sources
  • Input or output
  • LLM call
  • Code
  • Check or gate
  1. Question “Which region drove the Q1 variance, and what does the budget note say about it?”
  2. Plan The model decides which tools answer this. It writes no figure itself.
  3. Run SQL A query runs on the analyst’s own tables. The variance by region comes back exact.
  4. Search documents Hybrid retrieval over the budget notes: BM25 and vectors fused, then a cross-encoder keeps the best passages.
  5. Check Every figure traces to the query and every claim to a passage. If the evidence were thin, it would say so.
  6. Cited answer Two sentences, a bridge chart, and the page each claim came from.

Other routes through the graph

  • No numbers in the question: the plan skips SQL and searches the documents only.
  • Not enough evidence: back to the plan with a better query, a few rounds at most, then an answer that says what is missing.

All projects

Generative AI first: an LLM agent built at Natixis, then the RAG systems, agents and evaluation tools I build outside work, and two forecasting and full-stack products. Source code on GitHub.

At BPCE & Natixis

RAG and LLM systems

  • Sanad Agentic RAG over private documents where every claim resolves to a page and a bounding box. Case study

    Reads nine formats, including scans and photos through vision OCR; chunks on structure and never splits a table row; adds Contextual Retrieval. Hybrid BM25 + vector retrieval fused with RRF, then a cross-encoder rerank: on the sample corpus it lifts a weak embedder’s MRR from 0.53 to 0.94. A LangGraph agent rewrites follow-ups, decomposes multi-part questions, critiques its own evidence and declines rather than guessing. Every citation quote is verified against its source chunk, and an eval harness scores recall, MRR and groundedness per model. Ingestion pauses and resumes mid-file without paying for any call twice.

    Built with LangGraph, FastAPI, hybrid BM25 + vector retrieval (RRF), cross-encoder rerank, Claude vision OCR, ChromaDB, SQLite FTS5, pausable/resumable pipelines, Angular

  • Applied LLM Projects Monorepo of five LLM systems behind one auto-discovering gateway with a unified cost tracker.

    A function-calling SQL agent with guardrails, a RAG eval harness (recall@k, LLM-judge groundedness, chunking ablations), a forecasting narrator with rolling-origin backtests, a drift sentinel with small-to-large model triage, and a distillation pipeline measuring an accuracy-vs-cost Pareto frontier.

    Built with FastAPI gateway, OpenAI-compatible local models and all major LLM providers, RAG evals, distillation, cost tracking, Angular

LLM agents and evaluation

Forecasting and full-stack products

  • Vantage / Insight Terminal Full-stack market intelligence platform over ~194 instruments: equities, crypto and commodity futures.

    An ensemble of XGBoost, LightGBM, a PyTorch LSTM, ARIMA and Prophet against naive baselines with honest walk-forward backtesting, LLM-scored news sentiment, and an AI-written daily brief served from a local OpenAI-compatible model.

    Built with FastAPI, SQLModel, PyTorch LSTM, XGBoost/LightGBM, ARIMA/Prophet, walk-forward backtesting, Angular

  • Tahanot ERP Self-hosted, Odoo-inspired mini-ERP (sales, purchase, inventory, dashboard) rebuilt from scratch. Case study

    A generic model registry with domain filters, state-machine documents, quant-based stock with backorders, product variants, invoice payments and credit notes, plus a tool-calling AI assistant over the service layer. Bilingual EN/FR, with Postgres concurrency and locking tests.

    Built with FastAPI, SQLAlchemy, PostgreSQL, Alembic, Docker, tool-calling assistant, React

Skills

  • Generative AI & LLMs
  • RAG & AI agents
  • LLM evaluation
  • Prompt engineering
  • Machine learning
  • Computer vision
  • NLP
  • MLOps

Stack

GenAI & LLMs
LangGraph, LangChain, Hugging Face, LLM APIs (Claude, Gemini, OpenAI-compatible), ChromaDB, spaCy, NLTK
Code
Python, Scala, SQL, JavaScript
ML & deep learning
PyTorch, TensorFlow, Keras, JAX, Haiku, scikit-learn, XGBoost, LightGBM
AI apps & APIs
FastAPI, Chainlit, Streamlit, React, Angular
Data
Pandas, Polars, Apache Spark, Hadoop, DuckDB
Databases & search
SQL / NoSQL, PostgreSQL, Impala, MongoDB, Cassandra, Elasticsearch
Platforms
Docker, Kubernetes, OpenShift, Jenkins, GCP (Vertex AI), Dataiku

Methods

Generative AI
agents and tool calling, RAG with hybrid retrieval (BM25 + vectors + RRF), reranking, LLM evaluation and LLM-judge grading, prompt versioning and A/B testing
Machine learning
OCR, object detection (YOLO), speech recognition (Whisper), time-series forecasting
Engineering
ETL pipelines, CI/CD, Git, Agile (SAFe, Scrum)

Ask this CV

Type a question. A BM25 ranker, the keyword half of Sanad’s search, scores every passage on this site in your browser and answers with the passages themselves, so it can’t invent anything.

3 best of 52 passages, ranked by BM25

  1. Skills: Generative AI score 3.45

    agents and tool calling, RAG with hybrid retrieval (BM25 + vectors + RRF), reranking, LLM evaluation and LLM-judge grading, prompt versioning and A/B testing.

  2. Project: Insight Generator score 2.80

    LLM agent for finance spreadsheets: clean data, exact pivots, answers computed in SQL. Finance teams upload Excel or CSV exports; it sniffs the real format, parses French and English finance values, flags quality issues, versions every change and builds Excel-style pivots with exact subtotals. The …

  3. Project: Applied LLM Projects score 2.70

    Monorepo of five LLM systems behind one auto-discovering gateway with a unified cost tracker. A function-calling SQL agent with guardrails, a RAG eval harness (recall@k, LLM-judge groundedness, chunking ablations), a forecasting narrator with rolling-origin backtests, a drift sentinel with …

Try

About

I’m Hamza Ouhssaine. Tamazight is my native language, and I work in French, English and Arabic.

I trained in mathematics in MP* at Ibn Ghazi, then at Centrale Lille, with a master’s in applied mathematics from the University of Lille alongside.

I started in research, detecting data hidden in images at the CRIStAL lab, then built OCR and fraud-detection APIs at Meelo. At Amadeus I built the data pipelines airline revenue management runs on, and led hiring for the team.

At ArcelorMittal I put computer vision on the steel line, reading product IDs at a 0.098% character error rate.

Since 2026 I build GenAI for finance at BPCE & Natixis: agents that compute rather than guess, and answers that cite their sources.

Hamza Ouhssaine GenAI Engineer & Senior Data Scientist

Building something like that? Write to me.

hamzaouhssaine@gmail.com
The two leaning glass towers of Tours Duo in Paris, the taller one carrying the Groupe BPCE sign, above railway tracks.

Contact

Have a GenAI, RAG or machine-learning project in mind? Email is the quickest way to reach me.

Based in Paris, France. Open to relocation.

Languages
French, English, Arabic, Tamazight

Write to me about

Email me