Post by Andrei Dumitru

🧵 Architecting an Automated, GDPR-Compliant 4K Media Pipeline. I recently engineered an automated system capable of turning raw footage into fully GDPR-compliant, production-ready 4K assets. This project went far beyond basic scripting; it required designing robust data flows, optimizing model state management across frames, and ensuring seamless cross-platform adaptability across NVIDIA (CUDA) and AMD (ROCm) environments. The mission? Learn AI technologies by solving a practical problem, automate the identification and redaction of Personally Identifiable Information (PII), such as faces and license plates, to turn legal compliance into a frictionless, built-in feature of the creative workflow. To achieve this, I built a resilient, multi-stage pipeline bridging advanced computer vision with intensive video processing tools like FFmpeg and GLSL rendering. My modular approach includes: 🔍 Stage 1: Multi-Model Detection (detections.py): Initiates detection using multiple DL models (e.g., people, car > plate). A key technical challenge overcome was improving robustness via sophisticated logic for intra-frame duplicate checking and dynamic resolution scaling of models based on weight analysis. 💉 Stage 2: Interpolation & Filtering (interpolation.py & filters.py): To prevent data loss, an interpolation layer synthesizes missing PII. Redaction is managed by dedicated filters (like color-masked boxes), ensuring professional visual fidelity. ⚙️ Stage 3: Orchestration & Execution (one-shot.sh): A master control system manages quality settings and handles complex processing, culminating in final rendering using advanced filters (GLSL blur encoding) for guaranteed 4K HEVC output. The strength of the one-shot.sh script is its reliable sequence: Detection > Filter > Conversion. It proves expertise in building end-to-end industrial systems meeting demanding performance criteria across multiple hardware stacks. 🔥 Performance Highlights (on a AMD Radeon 890M): Processing Frames: 30+ FPS Processing Plates: 90+ FPS Binding Plates to Masks: 16,894.11/s Encoding: 84 FPS I'm incredibly proud of this demonstration project! I learned a lot across multiple technical areas like computer vision, media engineering, and scalable software architecture - all while keeping cross-platform deployment and high throughput at the forefront. Always excited by problems that require bridging advanced AI theory and reality. #AI #MachineLearning #ComputerVision #VideoProcessing #GDPRCompliance #SoftwareArchitecture #CUDA #ROCm