Luis A

Edge AI Engineer

VisionTech • Período integral

Aug 2025 - Present • 11 mos

Built an internal model optimization library used by multiple engineers to prepare deep learning models for edge deployment. The library supports: Channel-level structured pruning to reduce model size and increase throughput INT8 quantization to ONNX format for hardware-agnostic edge deployment Compatibility with TensorRT and ONNX Runtime inference frameworks Owned the end-to-end architecture of a high-throughput video processing pipeline combining Object Detection, Multi-Object Tracking, and CLIP-based semantic embeddings. Designed and implemented a video-text retrieval backend using Milvus and FastAPI, enabling sub-second semantic search across weeks of archived video footage. The system allows users to retrieve complex events such as abnormal behaviors or security incidents from large-scale historical indexes. Optimized inference pipelines for edge devices, achieving 180 aggregated FPS (6 streams @ 30 FPS) on a single NVIDIA Jetson Orin by migrating Python prototypes to a high-performance C++ implementation using ONNX Runtime with TensorRT execution providers. Reduced model size by approximately 95% (200MB → 10MB) by distilling CLIP image encoders into a lightweight MobileNetV3-Large architecture to meet strict memory constraints on edge hardware. Containerized the full edge deployment stack (video ingestion and search API) using Docker to ensure reliable production deployments. Trained and deployed production-grade object detection models including: YOLOv8 YOLOX YOLO-NAS Tech Stack: C++, Python, PyTorch, TensorRT, ONNX Runtime, Docker, Milvus, FastAPI

AI Engineer

dtLabs • Período integral

Mar 2023 - Aug 2025 • 2 yrs 5 mos

Prototyped vision-based logistics monitoring solutions for object tracking and anomaly detection in industrial environments. Developed high-performance prototyping tools in Python and C++ (FastAPI + OpenCV) to rapidly validate computer vision algorithms and deployment strategies. Built and deployed a face recognition and Automatic License Plate Recognition (ALPR) system using InsightFace and optimized C++ inference pipelines, including serverless components deployed on AWS Lambda. Engineered a robust OCR pipeline for automated text extraction in industrial environments, improving reliability under challenging lighting conditions through optimized image pre-processing and enhancement techniques. Reduced system latency and operational costs by migrating the video ingestion pipeline from AWS Kinesis to RabbitMQ, improving throughput and system responsiveness. Designed and implemented real-time monitoring dashboards (React) and integrated automated alerting for detection and recognition events. Trained and deployed object detection and multi-object tracking pipelines for a variety of industrial and logistics use cases. Tech Stack: C++, Python, OpenCV, FastAPI, React, AWS, RabbitMQ