Full data value chain coverage: From extraction to production.

Domain	Tools
Languages
AI & ML
Engineering
BI & Corp

I built a custom NAS (TrueNAS Scale) that evolved from family storage to a technical playground. I explore networking (VLANs, Reverse Proxy, VPN), hardware assembly, and run self-hosted services (Nextcloud, Media Servers, Pi-hole) via Docker.

Legend: ✅ = Completed (click title for repo) | 🚧 = Work In Progress

• ✅ Smart Outreach CRM

  • Solo • From scratch
  • Automates internship applications via LinkedIn parsing & GenAI with a Human-in-the-loop validation step.
  • Stack: Python Streamlit LLM APIs Docker Chrome Ext

• 🚧 Smart Shopping List Generator

  • Solo • From scratch
  • Digitizes physical recipe cards (HelloFresh) using OCR and GenAI for ingredient normalization/entity resolution.
  • Stack: Python Tesseract LLM Streamlit Docker

• ✅ clustVarACC (R Package)

  • Team of 3 • From scratch
  • Comprehensive R Library for Variable Clustering (K-Means, HAC, MCA) using R6 (OOP). Includes Vignettes & Unit Tests.
  • Stack: R R6 (OOP) Shiny Package Dev

• ✅ MedTriage-AI

  • Team of 4 • From scratch
  • AI Copilot for emergency medical triage based on the official FRENCH standard.
  • Dockerized multi-service architecture (Frontend, Backend, MLflow) deployed on Hugging Face Spaces.
  • Integrated a Pydantic-AI Agent with RAG capabilities (ChromaDB) for medical protocol analysis and structured data extraction.
  • Production-Grade & Ethical Focus: Prompt injection security, real-time FinOps (API costs) and GreenOps (EcoLogits) monitoring to track carbon footprint.
  • Stack: Pydantic-AI Mistral AI FastAPI Streamlit Docker MLflow ChromaDB EcoLogits

• ✅ Fraud Detection with Cost-Sensitive Learning

  • Team of 2 • Imbalanced Data
  • Fraud detection in highly imbalanced check transaction data (IR ~165:1, 4.6M transactions).
  • Dual Approach: Statistical optimization (F1-Score) vs Economic optimization (Profit maximization).
  • Achieved 93.3% profit capture rate (€2.14M) using Instance-Weighted XGBoost with custom cost matrix integration.
  • Stack: Python XGBoost Polars Scikit-learn Imbalanced-learn

• ✅ Electricity Load Forecasting & R Package

  • Solo • Time Series
  • Forecast building electricity consumption at 15-minute intervals using classical time series methods.
  • Benchmarked SARIMA, ETS, NNAR and implemented the Weighted Nearest Neighbors (WNN) algorithm as a native R package.
  • Stack: R forecast neuralnet Package Development

• ✅ Energy Performance Predictor (DPE)

  • Team of 4 • From scratch
  • Dual Model: Classification (Energy Class) & Regression (Consumption) served via API to a reactive frontend.
  • Stack: Python Shiny FastAPI Docker Scikit-learn

• 🚧 Job Market Insights & NLP

  • Team of 4 • WIP
  • Insights extraction from job descriptions using Topic Modeling (LDA) and semantic clustering.
  • Stack: Python NLP (Spacy/Gensim) Streamlit

LinkedIn • Email