Mexico
I am a Machine Learning professional with experience across academia and industry, focused on applied ML, computer vision, and LLM-powered systems. I hold an M.S. in Computer Science, where my thesis focused on thermal face recognition using CNNs. After completing my studies, I worked as a full-time university professor, teaching Artificial Neural Networks, Numerical Methods, and core computer science subjects. During this time, I authored and co-authored peer-reviewed papers on biometric facial recognition using deep learning and thermal imaging. In industry, I designed the first computer vision system in Mexico for processing national voter ID cards and extracting key regions of interest (names, addresses, photographs). The system combined YOLO, U-Net, and a custom data augmentation strategy, and was documented in the paper "ID Processing System Based on Deep Learning Models", published in Research in Computer Science and presented at COMIA. I have contributed to the open-source ecosystem through OCR systems, segmentation pipelines, and speech recognition models using architectures such as GANs, CRNNs, and U-Net. I also have experience with object detection from the R-CNN family (Fast R-CNN, Faster R-CNN) and instance segmentation with Mask R-CNN. Additionally, I have used YOLO for classification, semantic segmentation, and pose detection. I have experience building LLM-powered applications, including Agentic AI systems and Agentic RAG pipelines, using LangChain and LangGraph, with a focus on agent orchestration, tool integration, and structured multi-step workflows. I also have hands-on experience fine-tuning foundation models for text classification and instruction following, and optimizing the process with parameter-efficient methods such as LoRA In my current role as a Machine Learning Engineer, I design, deploy, and maintain large-scale ML pipelines in production, covering the full lifecycle from feature engineering and training to inference, monitoring, and CI/CD (AWS SageMaker, Airflow, Snowflake, Feast, Prometheus, Grafana). This combination of research depth, systems thinking, and production experience defines my approach to building intelligent systems that are both robust and scalable.
- Contributed to the open-source community by developing four deep learning-based reference kits showcasing performance optimization with Intel technologies such as Intel Extension for PyTorch, Intel OneDNN-optimized TensorFlow, Intel Neural Compressor, and Intel Distribution of OpenVINO: 1. OCR solution: automated text detection and extraction from documents using EasyOCR for text detection and a CRNN architecture for recognition. https://github.com/oneapi-src/historical-assets-document-process 2. Document classification system: extracted meaningful text from scanned claims and categorized them using OCR and CNN-based classification. https://github.com/oneapi-src/invoice-to-cash-automation 3. Automatic Speech Recognition (ASR): converted unlabeled speech into phonemized text with a GAN trained in an unsupervised manner. https://github.com/oneapi-src/ai-transcribe 4. Semantic segmentation solution: U-Net–based system to assist drones in safe landing by identifying and segmenting paved areas. https://github.com/oneapi-src/drone-navigation-inspection - Additionally, I developed a custom machine learning platform for large-scale deployment of diverse end-to-end pipelines, including Recommendation Systems, Sentiment Analysis, Real-Time Video Analytics, and Face Recognition. The platform was optimized for rapid debugging, hyperparameter tuning, distributed training/inference, and multi-architecture deployment. It also integrated a monitoring component with Grafana to track pipeline performance over time, leveraging tools such as Docker, Bash scripting, and GitHub Actions. - I also conducted a comparative performance analysis of a Document-Level Sentiment Analysis application across cutting-edge Intel CPU architectures, using both stock PyTorch and Intel Extension for PyTorch. The study included fine-tuning and inference modes, single vs. multi-instance deployments, multiple BERT models (including quantized versions), and diverse datasets.
- Building of the company’s first clinical Large Language Models (LLMs) in the form of a Question-Answering system using BioBERT and GPT2, which were implemented through Hugging Face based on TensorFlow/Keras and using a large dataset of about 30K instances. BioBERT produces the embeddings of each question and answer, then a Semantic Search module returns ranked question-answer pairs similar to the current question, and finally, GPT-2 generates the answer based on an input comprised by a concatenation of the current question with the ranked similar question-answer pairs. - Development of the company’s first automated Time Series system, which was built using Convolutional Recurrent Neural Networks through TensorFlow/Keras with the goal of Forecasting high blood glucose levels (hyperglycemia) in diabetic patients. The Recurrent Neural Network of the architecture uses Long Short-Term Memory (LSTM) cells. The model was trained and tested on a dataset of 12 simulated cases and 12 clinical cases. This dataset contains features like blood glucose levels, carbohydrate, and insulin data.
- Led the development of the first Mexican Computer Vision system for processing national voter ID cards, designed as an OCR pipeline. Built an end-to-end solution by cascading state-of-the-art models including YOLOv4, U-Net, RotNet, and a customized version of Keras-OCR, enabling the extraction of key information such as names, addresses, and photographs under uncontrolled conditions (e.g., rotations, lighting variations). To avoid overfitting during the training of YOLO, U-Net, and RotNet, I developed a data augmentation strategy implemented with NumPy and Pillow. The strategy increased the dataset size by applying techniques such as rotation and translation of the ID region, illumination changes, and pasting the IDs onto different random backgrounds. This data augmentation was applied offline. - Developed the first Mexican classification system for voter IDs using Convolutional Neural Networks, fine-tuning a VGG16-based architecture on segmented IDs and applying real-time augmentation during training using the data augmentation strategy previously mentioned. - Primary author of a research paper about the Computer Vision system for processing Mexican voter ID cards , presented at the Mexican Conference on Artificial Intelligence (COMIA) and published in the journal Research in Computer Science. - Co-author and principal technical contributor of a research paper on fingerprint recognition, accepted at the Mexican International Conference on Artificial Intelligence (MICAI) and published by Springer in the Lecture Notes in Artificial Intelligence (LNAI) series.
- Taught a comprehensive course on Artificial Neural Networks to 100 students, covering algorithms such as Densely Connected Neural Networks, Convolutional Neural Networks, and Recurrent Neural Networks. These algorithms were applied to the domains of Natural Language Processing, Computer Vision, and Time Series, with implementations in TensorFlow/Keras. - Taught Artificial Intelligence, Numerical Methods, Simulation, Operating Systems, Operating Systems Workshop, and Structured Programming to 180 students. - Authored and submitted two research papers on thermal face recognition using machine learning and deep learning techniques. Both papers were accepted: one was published in Systems and Computation journal, and the other in Innovation and Technological Development Digital Journal.