MY
PROJECTS
Object detection and 3D modeling
Development of an object detection and dimension estimation algorithm for storage. Objects are detected in images and their dimensions are estimated. They are then modeled in 3D to be integrated into a solution for optimizing storage space.
Anomaly Detection
Development of a real-time anomaly detection tool. This decision-support tool helps minimize labeling errors on automated production lines for seed bags. It detects labels on newly created bags and compares them to the correct labels. The project is divided into two phases: a label detection phase using a deep learning algorithm and a label comparison phase using image processing and machine learning algorithms.
Personal projects
Deep Learning with PyTorch
A set of deep learning projects using PyTorch. Fine-tuning of a Large Language Model (BERT), fine-tuning of multiple CNNs for classification tasks, model optimization through regularization and hyperparameter tuning, and implementation of gradient descent for a linear regression problem.
Knowledge-Based System (KBS)
Creation of an inference system supporting both binary and probabilistic logic. The knowledge base and rule set can be manually defined. Forward and backward chaining algorithms are available.
Combinatorial Optimization
Solving an NP-complete combinatorial optimization problem. The problem tackled is scheduling sports matches while adhering to strict constraints. A metaheuristic algorithm was implemented to find high-quality solutions within a reasonable time frame. I chose a local search algorithm using the descent method. To avoid stopping at the first local optimum encountered, I integrated multiple semi-random restarts that accept less optimal neighbors.
Document retrieval - NLP
This NLP project involves building a search system capable of retrieving relevant information from a specified question corpus based on a given document set. Key functionalities include implementing stemming to reduce words to their root forms and using a stop list to filter out common words with low informational value. The system supports queries of varying complexity, from basic unions or conjunctions of terms to advanced phrasal queries where terms must appear consecutively, complex logical expressions, wildcard usage, and tolerance for spelling errors. A user-friendly interface will allow users to view results, with query terms highlighted directly in the documents.
Reinforcement Learning
Development of a snake game with 0, 1, or 2 players. Reinforcement learning was used to improve the agents with each game played. The algorithms used include tabular Q-learning and approximate Q-learning. The A* algorithm was implemented to calculate distances.
Evolutionary algorithm
Implementation of an evolutionary algorithm for the One-Max problem. Three strategies were implemented: the steady-state algorithm, the estimation of distribution algorithm, and the compact genetic algorithm. An optimization was made by using an automatic selection of the mutation operator. Two methods were tested for this: adaptive roulette and UCB.
Machine Learning with Scikit Learn
Creation of a notebook illustrating the use of scikit-learn for classification, regression, and clustering tasks. Different datasets were used. This project is part of the online course: Building Machine Learning Solutions with scikit-learn.
Minizinc & Angular Project
Creation of a web interface offering four constraint satisfaction problems. Problem-solving is handled with MiniZinc, and the web interface is built with Angular. Users can also create their own problems by specifying variables and constraints. A system of constraint propagation and backpropagation enables dynamic display of the problem-solving process.