Travaux réalisés

Projets

Une sélection de projets mêlant géomatique, intelligence artificielle et développement full stack, du prototype de hackathon aux applications de terrain.

Featured mission

Water Credit AI

Mobile

Water Credit AI is a Flutter mobile app built in 48 hours during the UNESCO Water4Future Hackathon (February 2026, remote), designed as a water-credit marketplace for irrigated agriculture. The app offers two interfaces: a manager view (ORMVAG) tracking allocations, consumption, and transactions by basin, and a farmer view tracking personal water quotas, quick actions (irrigate, check the market, view history), and water-stress alerts. A built-in marketplace lets farmers trade unused water credits, with a real-time market price. The vision pitched to the jury went further: pairing this quota system with automated water-stress detection via Sentinel-2 imagery and NDVI indices, refined by machine learning models for irrigation recommendations — a direction explored in the pitch, but not implemented within the hackathon's timeframe. The project was presented to an international jury at the end of the 48 hours.

Flutter (Dart)

View details View demo

12 projects

Mobile

Zida Solaire Mobile App

Mobile app built for Zida Solaire, letting customers order directly from their phone. Users can browse the product catalog by category, place cash-on-delivery orders, create an account to request installation quotes, view company information and services, and contact the team via WhatsApp, phone, or email directly from the app.

Web

Jackson Energy Website

Business and e-commerce website for Jackson Energy, a Burkinabé company specialized in selling and installing solar, electrical, and electronic products. The site showcases the product catalog and company services, with a clear goal: improving online visibility and driving sales.

GeoAI

Land Cover Classification with CNN

Designing and training a convolutional neural network (CNN) with TensorFlow/Keras on the EuroSAT satellite dataset (Sentinel-2 imagery, 10 land cover classes: forest, industrial area, highway, residential area, crops...). After 12 training epochs, the model reached 81% accuracy on the test set, performing well on visually distinct classes (forest, industrial zones, water) with more confusion between visually similar classes (highways, permanent crops).

GeoAI

Predicting Healthcare Accessibility in Morocco

End-of-module GeoAI project: a Random Forest model predicting healthcare accessibility by municipality in Morocco, using geospatial features (distance to the nearest health center, population density, number of nearby facilities...). The first training run hit 100% accuracy, a classic sign of data leakage, which was identified and fixed, resulting in a realistic 80% accuracy model with a consistent confusion matrix across accessibility classes.

GeoAI

Querying a Flight Booking Database

Machine learnia vacation workbook: querying a flight-booking database in SQL (passengers, flights, bookings) to answer concrete business questions: SELECT queries, aggregations, joins, and subqueries.

GeoAI

Predicting the 2026 World Cup Winner

Machine Learnia vacation workbook: a supervised classification model (Random Forest) trained on 30+ years of international match results to predict football match outcomes, then used to simulate an entire tournament through to the 2026 World Cup final.

GeoAI

Tangier Land Cover Classification with Random Forest

Supervised land cover classification of Tangier using Random Forest, based on Sentinel-2 imagery (raw spectral bands plus NDVI, MNDWI, and NDBI indices) and ground-truth data, distinguishing four classes (agriculture, built-up areas, tree cover, water). Cross-validation score: 86%.

Web

Zida Solaire Website

Business and e-commerce website for Zida Solaire, a company specialized in selling solar, electrical, electronic, and home appliance products, as well as solar, electrical, and hybrid installations and electric well drilling. The site brings together the product catalog, service overview, and contact options.

GeoAI

Mapping the 2027 Tour de France with Clustering

Machine Learnia Vacation Workbook: grouping 120 villages into 21 stages using unsupervised clustering (K-Means, hierarchical clustering), then computing the optimal route for each stage with a greedy heuristic and the 2-opt algorithm (a variant of the traveling salesman problem). The analysis revealed that the provided village list produced stages far longer than the Tour de France's real average (over 200 km versus roughly 150 km), a concrete dataset limitation identified and documented.

GeoAI

Virtual Assistant for Hôtel Le Belvédère

Machine Learnia vacation workbook: building a virtual hotel front-desk assistant in three stages. First, a general-purpose LLM (Qwen2.5, open source, run locally) left to its own devices, caught hallucinating on things it didn't actually know. Then, the full hotel documentation added to the prompt, with a clear improvement but slower responses. Finally, a real RAG system (Retrieval-Augmented Generation, using sentence-transformers for embeddings) that only feeds the model relevant passages: accurate answers (hours, pricing, wifi), verifiable sources, and an honest refusal when a question falls outside the documentation.

Desktop

Afriki Achats — Cross-platform Purchase Management

Purchase and inventory management app for small businesses, built with Ionic Vue and Capacitor. It centralizes tracking of articles, suppliers, orders, and transport, with a real-time dashboard. Data storage automatically adapts to the platform (SQLite natively, localStorage on the web) for offline use. Available as a web app, Android app, and desktop app (Windows/macOS/Linux) via Electron — a single codebase powering all three targets.

Web

Burkina Solar Energy — Full-Stack Web Platform

Complete showcase website and admin backoffice for a solar energy company based in Ouagadougou, Burkina Faso. The application includes a product catalog (panels, inverters, batteries), a portfolio of completed installations, a contact form with message tracking, and an admin panel allowing full content management without touching the code. Decoupled architecture with a JWT-secured REST API, connected to a cloud-hosted MySQL database.