Scalable AI-Powered Media Processing & Metadata Generation System
2023 – Present · Personal Project
Scaling usability and streamlining image and data acquisition through open-source LLMs running on custom-built on-prem infrastructure.
Designed and deployed a comprehensive AI media pipeline capable of processing, tagging, and generating metadata for large-scale media collections. Built on AMD EPYC servers with GPU acceleration, this system handles 10TB+ of media assets using offline and cloud-hosted AI models.
The pipeline integrates LLM-driven tools for transcription, summarization, and auto-tagging while generating layered metadata formats compatible with Adobe Bridge and professional workflows.
The pipeline processes media through multiple AI stages, generating comprehensive metadata in TIFF, PSD, and JSON formats optimized for professional creative workflows.
Dual-socket servers with NVIDIA acceleration via NVLink
Virtualization & container orchestration
AI model inference & pipeline processing
Containerized AI services and LLM inference
High-performance storage architecture
Professional metadata integration
Identified the need to organize text, video, and image media in a system that could be automatically cataloged and retrieved based on semantic inputs.
Brought myself up to date on current components used in AI infrastructure — GPUs, PCIe generations, memory bandwidth, and containerization strategies.
First iteration built around available components: a SuperMicro H11DSi board with dual EPYC 7571 processors and a GeForce 1080 GPU running Proxmox. I rewrote fan controls to manage thermals with dual Noctua SP3 coolers, but PCIe 3 bandwidth proved limiting.
Rebuilt around a SuperMicro H13SSL board with a single EPYC 9334 CPU. The SP5 socket delivers better thermals, lower power draw, PCIe 5, and DDR5 memory support.
After testing several GPUs, settled on dual Quadro RTX 5000s connected via NVLink, providing 32GB of video RAM. Ollama runs in a Proxmox container; OpenWebUI VMs connect to the inference container.
Moved the SuperMicro tower into a full network rack alongside a Netgate SG-5100 pfSense firewall, managed switches, APC UPS, and supporting components.
Currently researching, testing, and building out the full media pipeline using open-source tools and locally running models.
Designed single-socket AMD EPYC systems with optimized thermal management and PCIe lane distribution for maximum GPU utilization.
Integrated multiple AI models for image analysis, video processing, and natural language processing with dynamic resource allocation.
Built comprehensive metadata generation supporting industry-standard formats with automated quality validation.
Seamless Adobe Bridge compatibility enabling direct integration into existing creative workflows without disruption.
This project demonstrates expertise in AI systems architecture, high-performance computing, and automated media processing pipelines.
Discuss Your Project View More Projects