6-month internship in Generali's Tech & Ops division. Built a Python script to automatically inventory all Power BI reports across the division (Streamlit MVP in one afternoon), deployed 15+ Power Automate workflows saving ~10 person-days/month, fully automated the REPERE training cycle end-to-end (HR ingestion → conditional emails → attendance tracking → PDF generation), rebuilt KPI dashboards with access controls and data validation, and delivered a production conversational AI agent (Copilot Studio) for project managers to query internal documentation in natural language.
Scraped and structured 2026 investment outlooks from 20+ major European asset managers (Amundi, BNP Paribas AM, Pictet, UBS AM, DWS, AXA IM…) managing >€10bn AUM. Built a full Flask web application : 10 pages including Executive Summary, Market Consensus, EdRAM Positioning, Commercial Angles, Peer Benchmark, and Geographic Lens : with interactive Plotly charts, URL-based filter sharing, and a premium ivory/gold design. Delivered to EdRAM's Head of Product Management for real sales strategy decisions.
Analyzed weather-to-incident correlation across SNCF's network. Engineered features (is_rush_hour, is_weekend, temperature delta, wind speed, season flags) and built a predictive model for daily default counts to enable proactive maintenance scheduling.
Ingested Board Game Arena's 1.92M player dataset, compressed to Parquet and queried with DuckDB for fast analytics. Segmented players into Casual / Regular / Hardcore clusters, built a bipartite player↔game graph and computed Jaccard similarity coefficients to map game affinity across the catalog.
Built a binary classifier for the French Ministry of Defense to detect AI-generated text vs. human-written content. Trained on a custom dataset with NLP preprocessing and fine-tuned classification architecture.
Analyzed BNP's internal ticketing system to identify SLA bottlenecks. Applied quantile regression to benchmark fair resolution times per category. Designed a smart task-reassignment engine to route overloaded teams to underutilized ones.
Analyzed Henkel's B2B hardware distribution revenue by store and region. Built geographic heatmaps of top-performing departments, modeled advertising ROI correlation, and produced concrete reallocation recommendations for commercial managers across Leroy Merlin and BricoMarché networks.
Modeled optimal product mix across Conserves, Bières and Petfood categories under hard store constraints (min. 5-store regional rollout before listing; max. 20% reference movement per period). Quantified cannibalization and substitution rates per SKU to maximize category revenue.
Audited 15,000+ rows of IT project management data: uncovered 640+ duplicates, 1,030 overlapping missions, 78% missing-data rate, and 31 contracts shorter than days worked. Designed GENERALI-A : an LLM-powered chatbot with Dynamic Query Reinforcement and per-project AI reliability scoring.
Modélisé le réseau maritime de CMA CGM (500+ navires, 420+ ports) comme un graphe pour calculer des routes optimales selon trois critères : rapidité, coût économique et émissions CO₂. Intégré un modèle météo/vent pour ajuster la vitesse réelle des navires, et proposé des solutions de décarbonation (kite Seawing, rotors Flettner).
Application d'investissement IA générant des recommandations de portefeuille personnalisées. Trois approches ML en parallèle : Random Forest/XGBoost pour la classification hausse/baisse, LSTM sur fenêtres 30 jours pour la prédiction de variation, et PPO (reinforcement learning) pour l'optimisation dynamique du portefeuille. Sentiment analysis VADER fusionné avec données Yahoo Finance et indicateurs macro (FRED).
3-week Business Deep Dive on LV's digital commerce with sales, customer cluster and product data. Identified South Korea and Middle East as priority growth vectors; designed AI-backed personalization framework, ultra-HNW client targeting model and 2026 sales forecasting architecture.
Data-driven media plan for a French outdoor equipment brand launching an eco-responsible product line. Allocated $1M across 5 ad channels (Meta, Snap, TikTok, DV360, Google Ads) using historical campaign data, conversion benchmarks and audience analysis.