← Back to portfolio

Skills map — AI Engineer Portfolio

Summary of projects, skills, evidence, soft skills and areas for improvement.

David MEDRAGH
Project 15 - Step 2

AI Engineer Portfolio - David MEDRAGH
  • 🧭 Projects & evidence
    • 👕 Project 2 - Serverless Hugging Face API for image segmentation
      • 🎯 Serverless access to the `segformer_b2_clothes` model to automatically segment images of clothes and people.
      • 🧩 Skills: computer vision, image segmentation, serverless AI API integration, secrets management (.env), pytest testing, mask visualisation
      • 🧰 Stack: Hugging Face Inference API (serverless), SegFormer B2 (segmentation vêtements/personnes), requests, Pillow, NumPy, matplotlib, python-dotenv, pytest
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet2, README.md, main.py, rapport_visualizations/, tests
    • 🏙️ Project 3 - Building energy consumption analysis
      • 🎯 Supervised modelling of CO2 emissions and energy consumption for buildings in Seattle.
      • 🧩 Skills: EDA, feature engineering, supervised regression, model comparison & selection, cross-validation, business interpretation of results
      • 🧰 Stack: pandas, NumPy, SciPy, scikit-learn (régressions, ensembles, GridSearchCV), seaborn, matplotlib, Jupyter
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet3, P3_01_nettoyage_analyse.ipynb, P3_02_modelisation_CO2.ipynb, P3_03_modelisation_energie.ipynb
    • 📊 Project 4 - Automatic classification of information
      • 🎯 Supervised classification to identify attrition causes at an IT consultancy, using HR, evaluation and survey data.
      • 🧩 Skills: supervised classification, class-imbalance handling, SHAP interpretability, multi-source cleaning (HR data, evaluations, survey), cross-validation
      • 🧰 Stack: scikit-learn, imbalanced-learn, SHAP, pandas, pyarrow, seaborn, PlantUML
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet4, projet4_etape1.ipynb through projet4_etape5.ipynb, data/*.csv, diagrammes PlantUML
    • 🚢 Project 5 - Deploying a Machine Learning model
      • 🎯 Turning an ML model into a production-ready service: API, CI/CD, database, tests and documentation.
      • 🧩 Skills: ML API design, model serialisation & persistence, CI/CD, containerisation, unit & functional testing, API documentation
      • 🧰 Stack: FastAPI, uvicorn, SQLAlchemy, PostgreSQL (psycopg2) / SQLite, joblib, Docker Compose, GitHub Actions, pytest + httpx
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet5, README.md, docker-compose.yml, docs/, artifacts/, tests
    • 🧪 Project 6 - MLOps, part 1
      • 🎯 First MLOps structuring around credit scoring: data prep, MLflow experiments, optimisation and business threshold.
      • 🧩 Skills: MLflow experiment tracking, gradient boosting (LightGBM, XGBoost), hyperparameter optimisation, decision threshold & business cost, data quality, secure serialisation
      • 🧰 Stack: MLflow Tracking, LightGBM, XGBoost, scikit-learn, skops, missingno, pyarrow/parquet, PostgreSQL
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet6, README.md, docker-compose.yml, mlflow_validation_screenshot/, data/*.parquet
    • 📡 Project 7 - MLOps, part 2
      • 🎯 Industrialising credit scoring: API, containerisation, CI/CD, monitoring, DEV/TEST/PROD environments and drift detection.
      • 🧩 Skills: multi-environment MLOps industrialisation, ONNX export & serving, real-time Prometheus/Grafana monitoring, automated data-drift detection, centralised prediction logs, E2E tests, performance profiling
      • 🧰 Stack: FastAPI + Pydantic Settings, ONNX Runtime (skl2onnx, onnxmltools), LightGBM, Evidently, Prometheus + Grafana Cloud, OpenSearch, Gradio, pytest + Playwright (E2E), Docker, GitHub Actions (DEV/TEST/PROD), Hugging Face Spaces, snakeviz + psutil (profiling)
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet7, README.md, .github/workflows/, Dockerfile, config/, doc/
    • 🔎 Project 8 - RAG system
      • 🎯 Designing and deploying a RAG system with OpenAgenda preprocessing, FAISS, LangChain, API and containerisation.
      • 🧩 Skills: end-to-end RAG architecture, embeddings & vector search, LangChain orchestration, LLM integration (Mistral AI), quantitative RAGAS evaluation, question-answering API
      • 🧰 Stack: LangChain (community, huggingface, mistralai), Mistral AI (LLM), sentence-transformers, FAISS, RAGAS, Hugging Face datasets, FastAPI, Docker, Postman
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet8, README.md, Postman/, Dockerfile, compose.yaml, doc/
    • 📐 Project 9 - Scoping an AI project
      • 🎯 Business scoping, planning, cost sizing, risk management and personal-data compliance.
      • 🧩 Skills: business scoping of an AI project, budget & ROI sizing, macro-planning & backlog, risk management, GDPR / DPIA compliance, system design
      • 🧰 Stack: templates de cadrage, system design, matrices de risques, NumPy + matplotlib (courbes budget)
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet9, README.md, livrables/, system_design/, templates/
    • 🩻 Project 10 - Semi-supervised approaches for image processing
      • 🎯 Exploring X-rays, feature extraction, clustering and semi-supervised learning.
      • 🧩 Skills: transfer learning (ResNet50 features), dimensionality reduction (PCA, t-SNE), K-Means clustering, semi-supervised learning (pseudo-labeling), medical image processing
      • 🧰 Stack: PyTorch, torchvision (ResNet50), OpenCV, scikit-learn (PCA, t-SNE, K-Means), Plotly, Pillow
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet10, projet10_etape1_exploration.ipynb through projet10_etape4_semi_supervise.ipynb, livrables/
    • 🎮 Project 11 - Reinforcement learning agent
      • 🎯 CartPole, FrozenLake, DQN exercises and the Eagle-1 mission on LunarLander.
      • 🧩 Skills: tabular Q-learning, DQN (replay buffer, target network), PPO via Stable-Baselines3, reward design & interpretation, TensorBoard training tracking, agent video reporting
      • 🧰 Stack: Gymnasium (CartPole, FrozenLake, LunarLander), PyTorch (DQN), Stable-Baselines3 (PPO), TensorBoard, MoviePy, Streamlit, FastAPI
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet11, README.md, notebooks mission, livrables/projet11_presentation.pptx
    • 🕸️ Project 12 - Multimodal data extraction from websites
      • 🎯 Local pipeline for source qualification, extraction, transformation, Airflow orchestration, MongoDB storage and KPI monitoring.
      • 🧩 Skills: orchestrated ETL pipelines, static & dynamic scraping, source qualification, NoSQL document storage, KPI monitoring, idempotence & error recovery
      • 🧰 Stack: Airflow (DAGs), MongoDB (pymongo), Playwright, BeautifulSoup4 + lxml, feedparser (RSS), NewsData API, Streamlit (dashboard KPI), Docker Compose
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet12, README.md, dags/, dashboard/, doc/uml/, livrables/
    • ♟️ Project 13 - AI agent
      • 🎯 An AI agent for learning chess: LangGraph, Milvus RAG, YouTube API, Angular and Docker packaging.
      • 🧩 Skills: tool-using agent design (tool calling), LangGraph orchestration, RAG on a Milvus vector store, external API integration (YouTube, Stockfish), full-stack AI development, Docker Compose packaging
      • 🧰 Stack: LangGraph, Milvus (pymilvus), sentence-transformers, Stockfish + python-chess, YouTube Data API, FastAPI, Angular 17, PyTorch, Docker Compose
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet13, README.md, backend/, frontend/, docker-compose.yml, livrables/rapport_technique.md
    • 🏥 Project 14 - Fine-tuning an LLM
      • 🎯 Medical-triage agent POC: dataset, SFT/LoRA, DPO, MLflow Registry, vLLM, FastAPI and monitoring.
      • 🧩 Skills: supervised fine-tuning (SFT + LoRA), preference alignment (DPO), GDPR anonymisation (Presidio + spaCy), high-performance LLM serving (vLLM), cloud GPU training, model registry & versioning, production LLM monitoring
      • 🧰 Stack: Qwen3-1.7B-Base, Hugging Face Transformers + Datasets + Hub, TRL (SFTTrainer, DPOTrainer), PEFT/LoRA, Accelerate, PyTorch fp16, vLLM (API OpenAI-compatible), spaCy + Presidio (anonymisation), MLflow Registry + DagsHub, Kaggle GPU T4, Lightning AI, Hugging Face Spaces GPU, FastAPI + Gradio, PostgreSQL (Aiven), Prometheus + Grafana Cloud
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet14, README.md, livrables/rapport_technique.md, doc/png/, doc/uml/, app/
    • ♾️ Project 15 - AI Engineer portfolio & closed-loop MLOps
      • 🎯 Personal technical project: drift detection, retraining, validation and automatic model promotion.
      • 🧩 Skills: closed-loop MLOps (drift → retrain → promotion), dual data-drift + concept-drift detection, AUC validation gate before promotion, self-hosting a monitoring tool, fail-fast CI/CD, tests & ≥ 90% coverage, dev → prod separation
      • 🧰 Stack: LightGBM + scikit-learn, Evidently 0.7 + Evidently UI self-hosted (Docker sur HF Space), River (ADWIN, DDM, PageHinkley), MLflow 3 + Model Registry (SQLite dev / DagsHub prod), GitHub Actions (7 jobs, crons quotidien + hebdo), pytest (93 tests, couverture ≥ 90 %), PlantUML (13 diagrammes), uv, python-docx + python-pptx (livrables générés)
      • 📎 Evidence: Dépôt GitHub du projet : https://github.com/davidmedragh/projet15, README.md, livrables/rapport_conduite_projet_P15.docx, scripts/, tests/, doc/uml/
  • 💼 Professional experience
    • Senior Developer / IT Innovation Advisor — Allianz Trade (Paris)
      • Trade credit insurance: the business domain behind my scoring projects (P6, P7, P15)
      • Technology watch and IT innovation advisory within the company
    • Tech Lead / Senior Developer — Groupe Zannier S.A.
      • Technical leadership and software development in the retail sector
    • Expert / Technical Lead, Java, JEE & BI — Société Générale (ART project)
      • Java/JEE, ETL and Business Intelligence expertise in the banking sector
    • JEE, Data & BI Engineer — RATP
      • Business Intelligence development: ETL, data warehouses and reporting
    • JEE, Data & BI Engineer — Monoprix
      • ETL pipelines and BI reporting in the retail sector
    • A data-engineering through-line
      • Several years in ETL, BI and data engineering — the direct foundation of my current ML and MLOps work
    • Computer Engineering degree — EPITA (2007-2009)
      • A senior profile: solid software-engineering and data experience predating my AI training
    • Cloud & AI certifications
      • Certified in multi-cloud AI: AWS, Azure, Microsoft, Oracle
      • AWS certifications: Cloud Practitioner, Enterprise Security, Cloud Architecture, Controlling Cost
    • Bridge between experience and training
      • My personal technical project (P15) closes the MLOps loop of a credit-scoring model — my employer's core business
      • Confidentiality respected: no internal project detail is disclosed
  • 🧠 AI stack covered
    • 📊 Data science & classic ML
      • pandas, NumPy, SciPy, matplotlib, seaborn, Plotly, Jupyter notebooks
      • scikit-learn: pipelines, preprocessing, cross-validation, GridSearchCV, RandomizedSearchCV
      • Models: linear regression, logistic regression, Random Forest, LightGBM, XGBoost, clustering, semi-supervised
      • Class imbalance: imbalanced-learn, resampling strategies and adapted metrics
      • Evaluation: AUC, F1, Average Precision, confusion matrices, business threshold, SHAP, data-leakage prevention
      • Quality & serialisation: missingno, skops, joblib, parquet/pyarrow
    • 🖼️ Deep learning & computer vision
      • PyTorch, torchvision and Hugging Face for deep-learning models
      • SegFormer B2 / image segmentation via the Hugging Face Inference API
      • ResNet50 transfer learning: feature extraction on X-rays
      • OpenCV, Pillow, PCA, t-SNE, K-Means, semi-supervised pseudo-labeling
      • Reading visual results, validation and business-facing reporting
    • 🎮 Reinforcement learning
      • Gymnasium: CartPole, FrozenLake, LunarLander (Eagle-1 mission)
      • Tabular Q-learning: Q-table, epsilon-greedy, exploration/exploitation
      • DQN in PyTorch: replay buffer, target network, reward curriculum
      • Stable-Baselines3 (PPO), TensorBoard tracking, MoviePy agent videos
    • 🔎 NLP, embeddings & RAG
      • sentence-transformers embeddings, vector search, chunking, retrieval and retrieval-augmented generation
      • FAISS and Milvus for vector databases
      • LangChain (community, huggingface, mistralai) for RAG orchestration
      • Mistral AI as the generation LLM, context-aware prompts
      • RAGAS / custom evaluation, Postman and a demo API
    • 🤖 AI agents
      • LangGraph for a tool-using agent (tool calling) and step orchestration
      • AI agent applied to learning chess: Stockfish + python-chess
      • Milvus RAG + YouTube Data API + FastAPI backend + Angular 17 interface
      • Docker Compose packaging and a full demo scenario
    • 🧠 LLM, fine-tuning & alignment
      • Hugging Face Transformers, Datasets, Hub, Qwen3-1.7B and tokenizer
      • PyTorch fp16, TRL SFTTrainer, TRL DPOTrainer, PEFT/LoRA, Accelerate
      • SFT, DPO, adapter merging, run tracking in MLflow
      • GDPR anonymisation of the corpora: Presidio + spaCy
      • vLLM, OpenAI-compatible API, Gradio, FastAPI, Kaggle T4 GPU, Lightning AI, Hugging Face Spaces GPU
    • ♾️ MLOps, monitoring & registry
      • MLflow Tracking, MLflow Model Registry, DagsHub, staging/production/archived lifecycle
      • Evidently for data drift (+ self-hosted Evidently UI in Docker), River for concept drift (ADWIN, DDM, PageHinkley)
      • Optimised ONNX export & serving: skl2onnx, onnxmltools, ONNX Runtime
      • Prometheus, Grafana, Grafana Cloud, OpenSearch, API/business/LLM metrics
      • GitHub Actions CI/CD, fail fast, secrets, quality gates, automatic promotion
      • Tests: pytest, coverage, httpx, Playwright E2E, smoke tests
    • 🏗️ Data engineering, API & cloud
      • Airflow, DAGs, multimodal ETL, MongoDB, parquet batches, snapshots and idempotence
      • Scraping: Playwright (dynamic), BeautifulSoup4 + lxml (static), feedparser RSS, NewsData API
      • FastAPI, OpenAPI/Swagger, /health /predict /metrics endpoints, token authentication
      • Docker, Docker Compose, Hugging Face Spaces, Aiven, Lightning AI, GitHub Actions
      • Angular, Gradio, Streamlit, dashboards and user-facing demo material
  • 🛠️ Technical skills
    • 📊 Classic ML & data science
      • Scientific Python & environment - 5/5 ●●●●●
        • Projects: Projects 2 to 15
        • Stack / evidence: Python 3.12, uv, Jupyter notebooks, pandas, NumPy, SciPy, matplotlib, seaborn, reproducible scripts and project packaging.
      • Data preparation & analysis - 5/5 ●●●●●
        • Projects: Projects 3, 4, 6, 10, 12, 14
        • Stack / evidence: Cleaning, joins, feature engineering, quality control, parquet/CSV/JSONL datasets, source qualification and corpus preparation.
      • Supervised machine learning - 5/5 ●●●●●
        • Projects: Projects 3, 4, 5, 6, 7, 15
        • Stack / evidence: scikit-learn, regression, classification, Random Forest, Logistic Regression, pipelines, cross-validation, GridSearchCV/RandomizedSearchCV.
      • Business modelling & scoring - 5/5 ●●●●●
        • Projects: Projects 6, 7, 15
        • Stack / evidence: LightGBM, credit scoring, business threshold, AUC, F1, Average Precision, baseline/challenger comparison and promotion gate.
      • ML evaluation & interpretability - 5/5 ●●●●●
        • Projects: Projects 3, 4, 6, 7, 15
        • Stack / evidence: Metrics suited to the need, confusion matrices, thresholds, feature importance, SHAP, data-leakage prevention and business-facing reading of results.
      • Deep learning & computer vision - 4/5 ●●●●○
        • Projects: Projects 2, 10
        • Stack / evidence: Hugging Face SegFormer, PyTorch, image processing, segmentation, X-rays, feature extraction, clustering and semi-supervised learning.
    • 🧠 LLM, NLP, RAG & agents
      • NLP, embeddings & RAG - 4/5 ●●●●○
        • Projects: Projects 8, 13
        • Stack / evidence: LangChain, FAISS, Milvus, embeddings, retrieval, RAG/RAGAS evaluation, question-answering API and Postman collections.
      • AI agents & LLM orchestration - 4/5 ●●●●○
        • Projects: Project 13
        • Stack / evidence: LangGraph, conversational agent, RAG tooling, YouTube API integration, FastAPI backend and Angular interface.
      • LLM fine-tuning & alignment - 4/5 ●●●●○
        • Projects: Project 14
        • Stack / evidence: Hugging Face Transformers, Qwen3, TRL SFTTrainer/DPOTrainer, PEFT/LoRA, Accelerate, JSONL datasets, MLflow tracking and adapter merging.
      • LLM serving & inference - 4/5 ●●●●○
        • Projects: Project 14
        • Stack / evidence: vLLM, OpenAI-compatible API, FastAPI, Gradio, deterministic stub, T4 GPU, Lightning AI and Hugging Face Spaces GPU.
    • ♾️ MLOps, quality & production
      • MLOps tracking & registry - 5/5 ●●●●●
        • Projects: Projects 6, 7, 14, 15
        • Stack / evidence: MLflow Tracking, MLflow Model Registry, DagsHub, run versioning, staging/production/archived and model traceability.
      • Monitoring, drift & observability - 5/5 ●●●●●
        • Projects: Projects 7, 14, 15
        • Stack / evidence: Evidently, River ADWIN/DDM/PageHinkley, Prometheus, Grafana, Grafana Cloud, dashboards, API/business/LLM metrics and potential alerting.
      • CI/CD, testing & software quality - 5/5 ●●●●●
        • Projects: Projects 5, 7, 12, 13, 14, 15
        • Stack / evidence: GitHub Actions, pytest, coverage, Ruff, fail fast, secrets, smoke tests, E2E tests, quality gates and manual/scheduled workflows.
      • Backend APIs & service contracts - 4/5 ●●●●○
        • Projects: Projects 5, 7, 8, 13, 14
        • Stack / evidence: FastAPI, OpenAPI/Swagger, /health /predict /metrics endpoints, token authentication, Postman and exchange contracts.
      • Data engineering & orchestration - 4/5 ●●●●○
        • Projects: Project 12, Project 15
        • Stack / evidence: Airflow, DAGs, MongoDB, multimodal ETL, parquet batches, drift snapshots, resynchronisation and idempotence.
      • Cloud, containers & deployment - 4/5 ●●●●○
        • Projects: Projects 2, 5, 7, 14, 15
        • Stack / evidence: Docker, Docker Compose, Hugging Face Spaces, DagsHub, GitHub Actions, Aiven, Lightning AI and secrets management.
    • 🧩 Product, architecture & communication
      • Frontend & demo interfaces - 3/5 ●●●○○
        • Projects: Projects 5, 13, 14
        • Stack / evidence: Angular, Gradio, dashboards, API documentation, demo screens and browsing ergonomics.
      • Technical communication & architecture - 4/5 ●●●●○
        • Projects: Projects 9, 12, 13, 14, 15
        • Stack / evidence: Detailed READMEs, technical reports, PowerPoint decks, PlantUML, C4, UML, flow diagrams and plain-language explanation.
      • Project management & AI scoping - 4/5 ●●●●○
        • Projects: Projects 1, 9, 15
        • Stack / evidence: Business scoping, needs, risks, ROI, planning, trade-offs, deliverables, evidence and reflective capacity.
      • Technical English - 5/5 ●●●●●
        • Projects: Documentation and AI ecosystem
        • Stack / evidence: Fluent English: professional exchanges, technical documentation, APIs, AI/LLM libraries, cloud platforms and integration guides.
  • ♾️ MLOps & industrialization
    • ML evaluation & interpretability - 5/5 ●●●●●
      • Projects: Projects 3, 4, 6, 7, 15
      • Stack / evidence: Metrics suited to the need, confusion matrices, thresholds, feature importance, SHAP, data-leakage prevention and business-facing reading of results.
    • AI agents & LLM orchestration - 4/5 ●●●●○
      • Projects: Project 13
      • Stack / evidence: LangGraph, conversational agent, RAG tooling, YouTube API integration, FastAPI backend and Angular interface.
    • MLOps tracking & registry - 5/5 ●●●●●
      • Projects: Projects 6, 7, 14, 15
      • Stack / evidence: MLflow Tracking, MLflow Model Registry, DagsHub, run versioning, staging/production/archived and model traceability.
    • Data engineering & orchestration - 4/5 ●●●●○
      • Projects: Project 12, Project 15
      • Stack / evidence: Airflow, DAGs, MongoDB, multimodal ETL, parquet batches, drift snapshots, resynchronisation and idempotence.
    • Cloud, containers & deployment - 4/5 ●●●●○
      • Projects: Projects 2, 5, 7, 14, 15
      • Stack / evidence: Docker, Docker Compose, Hugging Face Spaces, DagsHub, GitHub Actions, Aiven, Lightning AI and secrets management.
  • 🗣️ Languages
    • French - 5/5 ●●●●●
      • Native language — writing reports, presentations and defences.
    • English - 5/5 ●●●●●
      • Fluent — day-to-day professional exchanges, documentation, APIs, papers and the AI ecosystem.
  • 🤝 Soft skills
    • Analytical thinking
      • I connect business needs to metrics, risks and technical choices.
    • Problem-solving
      • I break a difficulty down into tests, hypotheses and reproducible validations.
    • Technical curiosity
      • I progressively broadened my scope from classic ML to RAG, agents, LLMs and MLOps.
    • Communication
      • I turn technical systems into readable material: READMEs, reports, UML, presentations and dashboards.
    • Autonomy
      • I can run a project end-to-end: scoping, implementation, testing, documentation and delivery.
    • Rigor
      • I favour evidence, metrics, tests and traceability over unverified claims.
  • 🪞 Reflective capacity
    • My view of the profession has evolved: I now see AI engineering as a full-stack skill across the information system's lifecycle, not just model training.
    • The MLOps projects taught me that a model's value depends as much on its monitoring, deployment and maintainability as on its initial metric.
    • In hindsight, I would structure evidence, screenshots and architecture decisions earlier to reduce the documentation effort at the end of a project.
    • I want to deepen advanced monitoring, multi-agent architectures and LLM evaluation under near-production conditions.
  • 🚀 Areas for improvement
    • Deepen advanced monitoring and large-scale model orchestration
    • Strengthen LLM and agentic-system evaluation
    • Structure project evidence earlier in the delivery cycle
    • Keep connecting technical choices to business impact
  • 🌐 Final portfolio
    • Final deliverable: step 3
    • Include the mind map, the personal technical project and the report
    • Take care of ergonomics, accessibility, navigation and visual evidence
    • Analyse market-demanded skills to guide what to learn next

Skill and language levels — diagram

Self-assessment illustrated with progress bars, grouped by domain (visual requirement of step 2).

Diagram of skill and language levels