Leopoldo López Reverón

I am a Machine Learning Engineer specializing in Computer Vision and Audio Intelligence, with experience developing end-to-end systems that encompass everything from dataset preparation to optimized inference and the creation of working demos I work with modern vision architectures (YOLO, U-Net, DPT, OCR pipelines) and spectrogram-based models for audio. I design and run ablation studies, model comparisons, advanced metric analysis (F1 macro/micro, mAP, precision/recall, logits, probabilities by class), and visualizations that allow understanding the internal behavior of the model I am particularly interested in building reproducible pipelines, optimizing models for production and integrating explainability techniques. My approach is practical, results-oriented, and based on a deep understanding of each stage of the model lifecycle

Generating a 360° panorama from a PTZ camera and estimating the depth of the entire scene

Both images can be enlarged by clicking to view reconstruction details

Biomedical segmentation project that detects and delineates cancerous regions in microscopic images The model generates precise masks that allow for the analysis of the extent and morphology of malignant areas

The images include: original sample and generated mask

Examples of the Personal Protective Equipment (PPE) detection system in different scenarios

The images show detection of helmets, vests and other safety equipment, with bounding boxes and real-time sorting

Pipeline End‑to‑End

The complete system includes all phases of the model lifecycle:

• Dataset: Collection, cleaning, and annotation of PPE images
• Train: YOLOv7 model with augmentations and validation
• Inference: Optimized script for real-time prediction

Ablation Studies

Model Comparisons

Visual comparisons of models

Final Results

2x2 comparison showing the original image, the super-resolution version and zooms of both.

A multilabel classification system capable of identifying multiple animal species from audio. Each clip is transformed into a Mel spectrogram and processed using a convolutional neural network. The model produces probabilities per class, logits, and advanced metrics such as F1, mAP, macro/micro accuracy, and recall.

GitHub

• Automated PPE compliance monitoring in industrial environments — Automation in Construction, 2025
• Maritime Surveillance by Multiple Data Fusion — VISIGRAPP 2023