Enzo – Python, PyTorch, OpenCV

A company specializing in religious events and celebrations needed an AI model that could analyze songs in real time, recognize them, and identify discrepancies with the musical genres and languages they wanted their guests to hear during the events.

• Designed and implemented the frontend, backend, machine learning models, and algorithms for a web application enabling real-time audio streaming from a user’s device.
• Built functionality for other users to listen live, with each stream processed in an isolated container to detect audio patterns such as loud bass, high tempo, and music style classification.

A sports company specializing in football wanted to develop a model to identify and track the ball's trajectory during free kicks.

• Developed a multi-camera system to reconstruct the 3D trajectory of a soccer ball during free kicks using synchronized video from four field-mounted cameras.
• Implemented ball detection on every frame, trajectory reconstruction algorithms, and automated event analysis to determine goals, impact points on the goal frame, and goalpost collisions.
• Provided interactive trajectory visualizations in JavaScript.

A company needed an AI model to detect brands in sports events. The system processes thousands of images, automating brand appearance tagging and enabling advanced search and analytics for sports organizations.

• Developed a computer vision pipeline to detect and classify over 200 brand logos in sports photography for a digital asset management platform.
• Combined OCR models with Siamese neural networks to match detected text and visual features against a limited training dataset.
• Designed a synthetic data generation process to expand training samples and improve model robustness.

A company wanted to build a swimming pool PoC for training.

• Developed an interactive Unity-based game designed for use in an endless swimming pool environment.
• Implemented real-time body and gesture tracking using OpenPose to enable gesture-based game control and to collect performance metrics (e.g., stroke count).
• Trained and integrated a custom YOLOv5 model to detect complex gestures (such as closed fists) that were not reliably recognized by standard pose estimation.
• Designed the system to operate in real time under challenging conditions (water movement, partial occlusion, and lighting variability).

A company aimed to develop a physics-based human motion estimation solution using monocular RGB video without the need for motion capture.

• Implemented the framework from Physics-based Human Motion Estimation and Synthesis from Videos, enabling physically plausible human motion synthesis without motion capture data.
• Developed a pipeline that refines noisy image-based pose estimates from monocular RGB videos by applying physics constraints via contact-invariant optimization

A near real-time warehouse floor monitoring project focused on detecting free and occupied areas to improve space utilization, operational efficiency, and layout planning.

• Built an application to measure the percentage of free warehouse floor space and support decisions on receiving new loads.
• Trained a semantic segmentation model in Detectron2 to classify pixels as free space or used space (e.g., boxes).
• Migrated the solution to edge deployment on NVIDIA Jetson Xavier, replacing Detectron2 with YOLACT-Edge for real-time inference.