ML & AI Projects
- Data lake
- Azure Functions
- Azure SQL + OpenAI catalogue
- AI Search
- Foundry agent + MCP
End-to-End Agent in Azure
Dataroots Academy graduation project: a correct, traceable and honest agent that searches structured and unstructured company documents, fully on Azure.
- •Smart ingestion: Azure OpenAI builds a semantic data catalogue
- •Vector search over documents with Azure AI Search
- •Custom MCP server on Azure containers for SQL search
- •Entra ID authentication
- Garmin tools
- LangChain agent + MCP
- SQLite
- Schedule + meal plan
IM Agent: AI Training Coach
A LangChain agent that reads your Garmin trainings and builds a training schedule towards your goal.
- •Custom tools to retrieve past trainings and metrics from Garmin
- •Considers weather and biometrics when planning
- •Also builds meal plans
- •Dockerfile and API ready for a Raspberry Pi
- Rolling windows
- MissForest imputation
- Feature selection
- XGBoost + LOAO CV
Master's Thesis: Hamstring Injury Risk
Predicting acute hamstring injuries in elite athletics and soccer from longitudinal monitoring data.
- •Severe class imbalance, handled with scale_pos_weight
- •Leave-One-Athlete-Out validation to avoid leakage
- •Generalisation tested across sports
- Orders
- Atomic paint jobs
- CP-SAT solver
- Schedule
Smart Scheduler for a Paint Shop
Constraint solver that schedules beams in an industrial paint shop, maximising on-time orders.
- •Colour changes, cleaning times, shift capacity
- •Live order data via a C# backend
- •Compared against an expert-made schedule
- TF-IDF
- Nearest neighbours
- Llama-3.2-1B
RAG for Culinary Knowledge
Naive RAG pipeline answering questions over a recipe database.
- •NLTK preprocessing
- •Evaluated with Precision, Recall, MAP
- Bi-Encoder top 100
- Cross-Encoder re-rank
- Top-k
Semantic Retrieval and Re-ranking
Two-stage retrieval that closes the vocabulary gap of sparse search.
- •Sentence-BERT + MS MARCO cross-encoder
- •Benchmarked on WikIR1k
- HAAR detection
- Eigenfaces + SIFT
- t-SNE
- RBF SVM
Crafting and Learning Features
Handcrafted versus learned features for face recognition in a low-data regime.
- •80 training images, 1,816 test images
- CNN classifier
- FCN / U-Net
- FGSM attack
In the Name of Deep Learning
Classification, semantic segmentation and adversarial robustness on PASCAL VOC.
- •20 object classes
- •Tested vulnerability to adversarial examples
- PRM planner
- SLSQP local planner
- EKF estimation
- Stanley control
Mobile Robot Navigation
Planning, control and state estimation for a robot avoiding obstacles in 2D.
- •Noisy GPS fused with control inputs
- •Differential drive and bicycle model
In progress
Reinforcement Learning
Project coming soon.
Thibault Willems
AI Engineer & Entrepreneur
MTEC Games
Immersive Experiences
ML & AI
Predictive Modeling & Data Science
Career
Front/Back-end
Scalable Architecture
Skills
Contact
Let's build something
Front & Back-end Projects
My Escaperoom
Booking platform for escape room owners, built with full ownership and end-to-end: architecture, payments, infrastructure, CI/CD and monitoring.
Tech Stack
OlavKTTT
Full-stack e-commerce platform for video creator Olav Koslosky. Includes a webshop for purchasing video content, a community forum with restricted access, and multi-tier user role management.
Tech Stack
MTEC Games
Frontend website for MTEC Games, a high-tech immersive escape room co-founded with my father. Showcases the experience with 3D WebGL visuals and interactive storytelling.
Tech Stack
Portfolio Thibault Willems
Personal portfolio showcasing past projects, skills and career journey. Built with a bento-grid layout, 3D models, and smooth Framer Motion animations.