Talvin Ackbaraly

Segmentation & Classification of Heritage Illustrations

Final-year research project in partnership with the Bibliothèque nationale de France (BNF). The goal is to automatically detect and classify illustrations (engravings, figures, ornaments) extracted from digitized heritage documents.

• Benchmarked object detection models: YOLO, Florence-2
• Defined and tracked evaluation metrics (mAP, IoU)
• Dataset management: annotation, cleaning, augmentation
• Weekly client meetings with BNF teams

Automatic License Plate Recognition

A complete license plate reading system (ALPR) developed in a team of two. Two approaches were compared: classic image processing and Machine Learning, with rigorous benchmarks.

• Classic approach: filters, morphology, segmentation
• ML approach: classification achieving 86% accuracy
• Comparative benchmarks Python vs C++ on Train/Val/Test datasets
• Full GitLab workflow: issues, merge requests, CI

Real-Time Video Motion Detection — GPU Port

Porting of a real-time video motion detection algorithm from CPU to GPU. The goal was to identify performance bottlenecks and maximize throughput using CUDA.

• CPU → GPU port achieving a ×24 performance gain
• Bottleneck analysis with Nsight
• Optimization of memory transfers and CUDA kernels

Medical Prediction & 3D Visualization

Prediction of Forced Vital Capacity (FVC) from pulmonary CT scans. The project covers the full pipeline: lung segmentation, feature extraction, ML prediction, and 3D visualization in a web interface.

• Lung segmentation and 3D reconstruction from CT scans
• FVC prediction using XGBoost
• Web interface with Next.js for results visualization
• Docker containerization and GitLab CI pipeline

Real-Time Fluid Simulation

A fluid simulation engine based on the SPH (Smoothed Particle Hydrodynamics) algorithm, rendered in real time with OpenGL. All physics computations are offloaded to the GPU via Compute Shaders.

• Real-time SPH simulation of thousands of particles
• GPU physics via Compute Shaders (SSBO)
• Rendering via Frame Buffer Objects (FBO)

Social Network Clone

Development of a full Twitter-like social platform in a team of 14, simulating real-world engineering conditions with rigorous DevOps practices.

• Microservices architecture: Java/Quarkus backend, React/Next.js frontend, MongoDB
• Features: authentication, timeline, likes, follows, real-time interactions
• Coordination across 14 developers with code reviews and project management

Full Compiler Construction

Implementation of a complete compiler for the Tiger language, covering all classical compilation passes from lexical analysis to type checking.

• Lexical and syntax analysis (Flex / Bison)
• Variable binding resolution
• Type inference and type checking

Digit Recognition & Sudoku Solver

A complete image processing pipeline: detecting a Sudoku grid in a photo, recognizing digits with a custom-trained neural network, and automatically solving the puzzle.

• Grid detection using classical image processing
• Custom CNN trained from scratch for digit recognition
• Integrated automatic Sudoku solver