About
AI Architect | Former Head of AI at ActiveEon | Ph.D. in Computer Vision & Machine Learning
20+ years in Software Engineering • 10+ years in AI, Computer Vision, Deep Learning, and Distributed Systems
Expert in AI at Scale, HPC+AI, MLOps, Computer Vision, Edge AI, Generative AI, and LLM Optimization.
I am an AI Architect and Researcher focused on bridging the gap between research innovation and industrial-scale deployment. My career combines deep technical expertise with strategic leadership, specializing in high-performance computing, computer vision, generative AI, and large-scale AI automation.
As Head of AI at ActiveEon (2017–2025), I led a team of Ph.D.-level engineers in designing large-scale AI infrastructures for industrial and research applications. My work focused on multi-GPU orchestration, distributed training, and cloud-edge hybrid pipelines, delivering mission-critical solutions for global clients.
With a software engineering foundation dating back to 2001 and specialized AI research since 2010, I excel in end-to-end development: from original algorithm design and GPU acceleration to rigorous optimization and production-ready inference across embedded, edge, and cloud environments.
Core Expertise
- AI Architecture & Scale — HPC for AI | Distributed & Parallel Workflows | Multi-GPU/node scaling | Federated & hybrid-cloud systems
- MLOps & Automation — Experiment tracking | Continuous Training (CT) pipelines | CI/CD for ML | AI workflow orchestration (cloud, on-prem, HPC) | Monitoring, drift detection, and model governance
- Computer Vision & Image/Video Processing — Object detection, anomaly detection, motion analysis | Real-time video analytics & CV pipelines | Background subtraction, low-rank/sparse decomposition
- Embedded & Edge AI — Real-Time Inference (Jetson, RPi) | TensorRT optimization | DeepStream/GStreamer | Constrained hardware inference
- AutoML — Hyperparameter Optimization (HPO) | Neural Architecture Search (NAS)
- LLMs & GenAI — Fine-tuning, PEFT/LoRA, Quantization | FlashAttention, AutoAWQ | Deployment (Triton, TGI, vLLM)
- Stack — C++, Python, CUDA, AI SDKs, APIs, JupyterLab Integration
Leadership @ ActiveEon — Former Head of AI (2017–2025)
At ActiveEon, I architected and delivered large-scale AI solutions tailored for industrial and research applications. My responsibilities included:
- Led a team of researchers in AI/ML, GenAI, and Vision Systems
- Architected ProActive AI Orchestration (PAIO) for scalable, cloud-native, and on-prem AI workflows
- Deployed LLMs, deep learning, and AutoML pipelines across multi-node, multi-GPU, and hybrid cloud environments
- Built custom SDKs and tools for dynamic workflow automation (e.g., ProActive Python SDK, ProActive Jupyter Kernel)
- Contributed to the ExtremeXP European research project by leading the implementation of a runtime for scheduling and executing complex analytics workflows on distributed infrastructure using ProActive AI Orchestration (PAIO). This includes integration for distributed AutoML, resource monitoring, and dynamic service orchestration via the ProActive Python SDK, enabling on-the-fly deployment of tools like TensorBoard and MLOps dashboards.
Academic Background & Research
Ph.D. in Computer Vision & Machine Learning – Université de La Rochelle, France
Specialized in:
- Low-rank & sparse matrix/tensor decomposition
- Subspace learning & optimization
- Multimodal and real-time video analytics
- Embedded computer vision (Jetson, Raspberry Pi, PandaBoard)
My academic research has led to 2800+ citations (h-index: 17) and contributions to peer-reviewed journals and top conferences (IEEE, Elsevier, Springer, CVIU, TIP, TNNLS, ICCV). I also actively contribute as a peer reviewer and open-source developer.
Open-Source & Contributions
🔗 GitHub: github.com/andrewssobral
- BGSLibrary (C++) – Widely used background-subtraction library for moving-object detection.
- LRSLibrary (MATLAB) – Framework for low-rank/sparse decomposition.
- MTT (MATLAB) – Tools for tensor manipulation and decomposition.
- OSTD (MATLAB) – Online stochastic tensor decomposition for multispectral video.
- VDTC (C++ & Python) – Vehicle detection, tracking & counting pipeline.
- GoDec (Python) – Low-rank + sparse decomposition.
- DTT (C++) – Header-only library for seamless data type conversions (Eigen, OpenCV, Armadillo, LibTorch, ArrayFire).
These projects are used in academia, industry, robotics, surveillance systems, and research labs worldwide.