I build AI that works in the real world — industrial floors, maritime decks, and edge-constrained hardware where latency, reliability, and observability are non-negotiable.

I currently work as a ML Engineer, designing end-to-end AI systems, guiding cross-functional teams, and taking applied research through to production deployment.

My core focus areas:

• Edge AI & Real-Time Computer Vision — Deploying optimized models (TensorRT, ONNX, CUDA) on constrained hardware with strict latency budgets
• Model Optimization — Quantization, pruning, distillation, GPU acceleration, and streaming pipelines
• Industrial & Maritime AI — PPE detection, depth estimation, multi-object tracking, and perception for complex real-world environments
• Systems Architecture — End-to-end system design, observability stacks, and production-grade engineering

Automated PPE Compliance Monitoring in Industrial Environments
Leopoldo López et al. — Automation in Construction, 2025, Elsevier

Real-time multi-modal perception system for PPE detection deployed across 6 production environments running 24/7. First-author publication in a Q1 journal.

Key results:

• +5 F1 points over state of the art on a custom industrial dataset
• 10× inference speedup via ONNX graph surgery → post-training quantization → TensorRT (120 ms → 12 ms @ batch=1, Jetson TX2, FP16)
• Validated over 6 months of real-world data across industrial and maritime environments

Stack: InternImage · YOLOv7 · OpenPose · TensorRT · ONNX · Jetson TX2 · PyTorch DDP

link

PyTorch · TensorFlow · YOLOv7/10 · InternImage · Mask2Former · OpenPose · MiDaS · MixNet
Detectron2 · MMDetection · RetinaFace · SegFormer · DPT / Vision Transformers
Object Detection · Semantic Segmentation · Pose Estimation · Monocular Depth Estimation · Multi-Object Tracking · Face Detection

TensorRT · ONNX Runtime · CUDA Kernels
Quantization · Pruning · Knowledge Distillation · GPU Acceleration · Latency Profiling

FFMPEG · NVENC · WebRTC · MinIO · Prometheus · ZFS
Real-Time Streaming Pipelines · Multi-Camera Systems · Observability · Reliability Engineering

Python · C++ · C · Java · MATLAB · Bash

Data Structures (Linked List, BST, B-Tree, Sorted Array · Java)
Numerical Methods (Gaussian elimination, LU/QR/Cholesky decomposition, iterative solvers, 2D convolution)
Optimization Algorithms (Simulated Annealing, Genetic Algorithms, Christofides, Constraint Programming via MiniZinc)
Compiler Design (Flex/Bison, LALR(1), symbol tables, bytecode generation)

Processing · OpenGL / GLSL · Custom toon/cel shaders · 3D surface rendering · Matplotlib · Animated depth map reconstruction

Monocular depth estimation on out-of-distribution domains (satellite imagery, lunar surfaces, PTZ panoramas)
Cellular automaton simulation · Julia Set visualization · Numerical PDE solving (FreeFEM++)

Production-grade safety system deployed in industrial and maritime environments.

• Multi-camera, multi-stream real-time inference engine
• TensorRT-optimized models for edge hardware
• Robust tracking, alerting, and full observability stack
• Designed to operate reliably under harsh, variable conditions

A research pipeline applying Dense Prediction Transformer (DPT) models to unconventional, out-of-distribution domains.

• Tiling strategy for high-resolution inference beyond model native resolution
• Crop alignment and seam blending for artifact-free stitching
• Multi-scale stacking for progressive depth refinement
• 3D surface reconstruction and animated GIF walkthroughs
• Domains: ESA Sentinel-2 coastal imagery, high-res lunar photography, PTZ panoramas

Evaluates 12 face detection methods (classical + deep learning) under adverse real-world conditions.

• Conditions: fog, rain, snow, night, sandstorm, crowded scenes
• Methods: Haar/LBP Cascades, Viola-Jones, RetinaFace, YOLO, Faster R-CNN, SSD MobileNet, RFCN
• Custom accuracy and inference-time metrics pipeline

Benchmarks 130+ pre-trained object detectors (Detectron2 + MMDetection) for beach litter detection.

• Cross-dataset class mapping: COCO/LVIS → TACO waste super-categories
• Full evaluation pipeline: IoU matching, precision/recall/F1 per class and per image
• Data augmentation (fog, rain, noise, flip) for YOLO-format training sets
• LaTeX table export for academic reporting

Five combinatorial optimization approaches to the Travelling Salesman Problem, benchmarked against Kaggle datasets up to 85,900 cities.

• Christofides approximation algorithm (≤1.5× optimal)
• Simulated Annealing with greedy initialization and 2-opt refinement
• Genetic Algorithm with custom crossover operators
• Angular sweep geometric heuristic
• Exact solver via MiniZinc constraint programming

A complete compiler for a statically-typed, C-like language, built from scratch.

• Lexer: Flex with type-aware identifier resolution
• Parser: Bison LALR(1) with single-pass syntax analysis, semantic checking, and Q-VM bytecode generation
• Symbol table, local/global scoping, full control flow (if/else, for, while, break, continue)
• Supports functions, recursion, integer, float, bool, string types

Production-style numerical computing library implementing:

• Linear solvers: Gaussian elimination, LU, QR, Cholesky decomposition
• Iterative methods: Jacobi, Gauss-Seidel, SOR
• 1D/2D cubic spline interpolation
• 2D convolution with boundary handling
• Integration with FreeFEM++ for PDE solving (MNIC project)

Master's in Smart Systems and Numerical Applications in Engineering (SIANI)
Universidad de Las Palmas de Gran Canaria (ULPGC) - CTIM Research Group (IEEE partner)

Bachelor's in Computer Engineering
Universidad de Las Palmas de Gran Canaria (ULPGC) — ranked top 5% globally in computer science (QS 2024)

I care about elegant, physically correct solutions, measurable impact, and systems that survive the real world.
I enjoy technical leadership, mentoring engineers, and translating complex research into clear, deployable engineering.

If you're working on applied AI — perception, edge deployment, optimization, or real-time systems — let's talk.

📩 ll11ll1@outlook.es · LinkedIn · Portfolio · GoogleSchoolar