Building a scalable question bank extractor using vision language models. Qwen2-7B on serverless GPUs, PyMuPDF text extraction, structured JSON output, and parallel batch processing of 1000+ question papers. The hard parts, the broken parts, and what I’d do differently.
I wanted my own LLM inference setup on bare metal — GPU-backed K8s, autoscaling that actually works with vLLM metrics, and auth so random internet people can’t drain my GPU credits. Here’s how it went.
I spent the weekend comparing IBM’s Docling and the Qwen2 Vision Model for document extraction. One is a structural architect, the other is a creative genius. Here is which one you should actually use.
Building a unified, multi-data-center OpenStack platform for CPU-based LLM inference and cloud services.