Available immediately — open to permanent roles

Samir Mahamat Saleh

AI Engineer · Data Scientist & Data Engineer

I build generative-AI and data systems that real teams actually use — and I measure, rigorously, whether they work. Two years doing exactly that inside the French Civil Aviation Authority.

  • Île-de-France, France
  • Permanent roles · hybrid or on-site
  • Microsoft Certified: Azure AI
2 yrs Apprenticeship at the French Civil Aviation Authority (DGAC/DSAC)
48 Structured JSON fields extracted from free-text safety reports
>80% Processing time saved per report — 30+ min down to about 5
1 Scientific preprint on LLM evaluation (SSRN, 2026)
About

Measurement over intuition

A dual profile: shipping end-to-end data and GenAI systems, and proving they hold up.

I hold a five-year degree (Bac+5, RNCP Level 7) in Artificial Intelligence and Data Science from aivancity School of AI & Data, I am Microsoft Certified: Azure AI Apps and Agents Developer Associate, and I am the author of a scientific preprint on LLM evaluation.

I spent two years as an apprentice at the DGAC/DSAC, the French Civil Aviation Authority, designing and delivering generative-AI tools used by real business teams in a sensitive, heavily regulated public-sector environment. Shipping to actual users — with their feedback, their edge cases and their compliance constraints — taught me more than any benchmark.

My work has two entry points. On one side, end-to-end data processing: multi-source ingestion, parsing, cleaning, quality and consistency checks, APIs and reporting. On the other, the design and above all the rigorous evaluation of RAG systems, agents and classification models — groundedness, citation faithfulness, calibration, and the discipline to publish a negative result when that is what the data says.

I am starting an MSc in Data Engineering & Cloud Computing in November 2026 to deepen the distributed-systems side: Spark, Kafka, Databricks and multi-cloud.

Skills

Technical toolkit

What I use day to day, grouped the way I actually work.

Languages

  • Python
  • SQL
  • TypeScript

GenAI, LLMs & agents

  • Dense & hybrid RAG
  • Embeddings
  • Chunking (parent-child)
  • FAISS
  • Azure AI Search
  • BM25 / TF-IDF
  • Reciprocal rank fusion
  • Reranking
  • Prompt engineering
  • Structured JSON output
  • Agents & orchestration
  • Self-consistency
  • MCP
  • LoRA / PEFT

Evaluation & reliability

  • RAGAS
  • Gold question sets
  • Ungrounded-answer rate
  • Citation faithfulness
  • Calibration
  • F0.5
  • Precision–recall
  • Bootstrap & Wilson CIs
  • Imbalanced data
  • Prompt-injection tests
  • Langfuse

Machine learning & NLP

  • Scikit-learn
  • TensorFlow
  • PyTorch
  • Classification
  • Information extraction
  • OCR
  • GANs (Pix2Pix)
  • SSIM / PSNR
  • Hugging Face
  • Qwen via Ollama

Data engineering

  • Pandas
  • NumPy
  • DuckDB
  • Multi-source ingestion
  • XML / PDF / JSON parsing
  • Cleaning
  • Data quality
  • Batch pipelines
  • Automation

Backend & interfaces

  • FastAPI
  • Flask
  • REST APIs
  • Streamlit
  • Gradio
  • PyQt5
  • React 18 + Vite
  • Tailwind CSS
  • React Flow
  • Recharts

Cloud & DevOps

  • Azure
  • Azure OpenAI
  • Azure AI Foundry
  • Docker
  • Git / GitHub
  • GitHub Actions
  • Azure DevOps
  • Tests & logging
  • Copilot · Cursor · Claude Code

Analysis & reporting

  • Analytical SQL
  • Power BI
  • Tableau
  • Matplotlib
  • Excel
  • Technical documentation
  • Explaining to non-technical teams

Governance & compliance

  • GDPR
  • EU AI Act
  • Responsible AI
  • Data confidentiality
  • Public-sector regulation

Currently learning — MSc 2026–2027

  • Spark
  • Kafka
  • MongoDB
  • Databricks
  • AWS
  • Terraform
  • Distributed algorithms
Experience

Professional experience

Five years of applied work, from start-up data analysis to GenAI in a regulated public authority.

  1. Data Science Officer — DGAC / DSAC Oct 2024 – Sep 2026

    French Civil Aviation Authority, Civil Aviation Safety Directorate Paris, France

    • Designed an AI solution for processing air-safety occurrence reports, turning free-text accounts into 48 structured JSON fields, with automated cleaning, integrity validation and consistency checks.
    • Automated the pre-filling of regulatory PDF forms from the extracted data, cutting processing time per report from over 30 minutes to about 5 — more than 80% saved on data entry and lookup.
    • Built the full application chain in Python: LLM extraction, JSON cleaning and validation, form generation, a FastAPI backend with several REST endpoints, and a web interface for business users.
    • Evaluated the solution on around a hundred real air-safety occurrences, verifying structured outputs, regulatory fields and data consistency before use.
    • Rolled it out to 5 business users with hands-on testing, feedback collection, fixes, technical and functional documentation, and iterative improvement.
  2. Python Developer Intern — UBA Chad Jun – Aug 2023

    United Bank for Africa — multi-branch banking data processing Chad

    • Built a PyQt5 desktop application letting business teams import, clean and consolidate Excel files coming from several bank branches.
    • Automated the data-cleaning steps with Python and Pandas, then generated Matplotlib visualisations to support analysis and reporting.
  3. Junior Data Analyst — AI Fashion Jan–Mar 2022 · Jul–Aug 2022

    Start-up project and year-validation internship, aivancity AI Clinic Remote

    • Collected and extracted web data from e-commerce sites and perfume and clothing brands, luxury and streetwear segments included, to feed product analysis.
    • Cleaned, structured and stored the datasets with Excel and Python/Pandas, then produced visualisations and actionable findings.
    • Built an olfactory benchmark based on ingredients and notes, with comparative analysis of fragrance components.
  4. Junior Data Analyst — Bunqr Sep – Dec 2021

    Start-up project, aivancity AI Clinic Remote

    • Accounting-automation project: reading invoices with OCR and algorithmically classifying documents into a chart of accounts.
Projects

Selected projects

Most of these are public on GitHub, with measured results rather than demos.

Final-year thesis

Where RAG Breaks — aviation-regulation QA

2025–2026

Citation-grounded question answering over EU aviation regulations, built to find out where conventional RAG fails and whether a calibrated escalation path fixes it: hierarchy-aware retrieval, hybrid search, reranking, verification and human review. 109 legal parents, 1,535 retrievable children, a 99-question evidence-linked benchmark with zero parent leakage.

  • RAG
  • FAISS
  • BM25 + RRF
  • Reranking
  • LoRA / PEFT
  • LaTeX
Open source

OpenResearch Scout Agent

2026

A tool-using research agent served over FastAPI: give it a topic and it searches arXiv, ranks the results and returns a source-grounded brief. Three tools (search, ranker, cost estimator), Pydantic schemas on both ends, guardrails covering prompt injection and credential-exfiltration patterns, and a regression eval harness scoring quality, safety, latency, schema validity and tool usage. Requests are logged to JSONL behind a /metrics endpoint; the Dockerfile targets Hugging Face Spaces.

  • FastAPI
  • Agents & tools
  • Pydantic
  • Guardrails
  • Eval harness
  • Docker
Open source

AI Act Evidence Desk

2026

An agentic RAG assistant that turns an AI use case into a source-grounded EU AI Act evidence map, for product owners and compliance analysts. Parent-child chunking, hybrid BM25/TF-IDF retrieval fused by reciprocal rank fusion, reranking, then three-pass synthesis with self-consistency (k=3) and an explicit critique step. Shipped with a three-tool MCP server, Langfuse tracing, versioned RAGAS results and five prompt-injection test scenarios.

  • Agentic RAG
  • MCP
  • RAGAS
  • Langfuse
  • Gradio
  • Python
Publication

VCBench — LLM evaluation on highly imbalanced data

2026

Companion code to an SSRN preprint: evaluating locally run open-weight LLMs (Qwen3-32B via Ollama, quantised GGUF) on binary classification with a very low positive rate. Compares vanilla, chain-of-thought, few-shot and hybrid prompting with F0.5, precision, recall, positive-prediction rate, PR curves, bootstrap and Wilson intervals, against a TF-IDF logistic baseline. The documented conclusion is a negative one: small-sample F0.5 gains are unstable without calibration diagnostics, and the classical baseline still wins.

  • LLM evaluation
  • Qwen3 / Ollama
  • Calibration
  • F0.5
  • Bootstrap CIs

Hospital Digital Twin

2026

A SimPy discrete-event twin of a hospital — ER, ICU and ward — with realistic boarding, where a full ICU blocks ER beds and congestion cascades backwards. A gradient-boosting model forecasts peak ICU load 24 hours ahead, backtested out-of-sample against a naive baseline. Under an epidemic wave, a targeted capacity package takes the median ER wait from 4,613 to 71 minutes and treats 63% more patients; adding ER beds instead would have matched the headline number while tripling boarding time. Two interfaces on one engine: a Flask/ECharts command centre and a Streamlit dashboard.

  • SimPy
  • Gradient boosting
  • Flask
  • ECharts
  • Streamlit
  • Backtesting
Open source

Multimodal Medical QA with BiomedCLIP

2026

Zero-shot medical visual question answering on WorldMedQA/V. BiomedCLIP is a CLIP-style model, so it cannot generate an answer — the task is posed as retrieval instead: every answer option becomes a candidate prompt, each is scored against the image, and the highest similarity wins. The app shows the score for every option rather than just the winner, which makes the failure mode visible where distractors look like the right answer. Includes a sampled batch evaluation against ground truth.

  • BiomedCLIP
  • Vision-language
  • Streamlit
  • PyTorch
  • Zero-shot
Open source

Smart Reading Interface

2026

A hands-free reader: words highlight as you say them aloud, and MediaPipe hand gestures drive zoom and page turns without touching the keyboard. The gesture classifier is geometric rather than learned — pinch and spread from thumb-to-index distance normalised by hand size, swipes from centroid velocity, with a cooldown so a held gesture does not fire repeatedly. Inspectable rules beat a black box when the model is steering someone's reading interface.

  • MediaPipe
  • PySide6
  • OpenCV
  • Speech recognition
  • Human-machine interface

YAKOOTA — generative-AI assistant

2024

An end-to-end document chatbot built on document parsing, embeddings and a contextual retrieval workflow, with robustness testing and successive iterations on answer quality.

  • Azure OpenAI
  • LangChain
  • Embeddings
  • RAG

Private project — details on request.

EuroMotion Model Factory

2026

A system-dynamics web app for strategic planning of an electric and software-defined vehicle supply chain: a JSON-driven simulation engine, interactive node diagrams and charted scenario outputs.

  • React 18
  • TypeScript
  • Vite
  • React Flow
  • Recharts

Cartoonizer GAN

2024

A Pix2Pix conditional GAN turning photographs into cartoon-style illustrations: full training pipeline, a shipped pretrained model, a PyQt5 desktop app and quantitative evaluation with SSIM and PSNR.

  • TensorFlow 2
  • Pix2Pix
  • OpenCV
  • PyQt5
  • SSIM / PSNR
Credentials

Certification & publication

Both verifiable online.

Microsoft Certified: Azure AI Apps and Agents Developer Associate

Microsoft · July 2026 — valid to July 2027

Scope: designing AI applications and agents on Azure AI Foundry and Azure OpenAI, RAG, orchestration, evaluation and responsible deployment.

Credential ID 3F9FED23D318BD9C Certification D6K6AD-CB45D0

Preprint — VCBench: evaluating open-weight LLMs

SSRN · 2026

A methodology study of the failure modes and calibration of locally run open-weight LLMs for rare-event prediction. An intentional negative result, documented with calibration diagnostics — because an evaluation that only reports wins is not an evaluation.

ssrn.com/abstract=6832218

Education

Education

  1. MSc Data Engineering & Cloud Computing — aivancity Nov 2026 – 2027

    RNCP Level 7: Expert in data infrastructure and large-scale data processing · taught in English Villejuif, France

    • Spark, Kafka, MongoDB, Databricks, AWS and Azure, distributed algorithms, data-oriented project management, cloud certifications.
  2. Programme Grande École — aivancity School of AI & Data 2021 – 2026

    Master-level degree in Artificial Intelligence and Data Science (RNCP38584, Level 7, Bac+5) Paris, France

    • A hybrid curriculum combining artificial intelligence, data science, business literacy, AI ethics, GDPR and the EU AI Act, project management, innovation and digital transformation.
    • Projects delivered through the AI Clinic, aivancity's real-world programme working on live needs from companies and start-ups.

Languages

French Fluent, professional
Native
English Master's taught in English · daily technical reading
C1

How I work

  • Methodical, and partial to measurement over intuition
  • Autonomous, with technical curiosity
  • Able to explain technical work to non-technical stakeholders
  • Systematic documentation
  • At ease in multidisciplinary teams
Contact

Let's talk

I am available immediately and open to permanent roles in AI engineering, data science and data engineering across the Paris region. If my profile fits something you are hiring for, I would be glad to hear about it.

Or write directly to