"Team Argentina team on AMD AI League Hackathon"

This project provides two pipelines built for the AMD AI League Challenges, running the Qwen2-VL-2B model on the ROCm ecosystem. The first is an OCR module designed to extract text from difficult images, such as angled license plates, blurry photos, and low-light street signs. The second is an Enterprise RAG engine that uses a local client-server setup and SQLite FTS5 to search through mixed-format corporate documents. It returns specific answers and exact file citations based only on the provided corpus. To ensure stability and stay within the 48 GB VRAM and 30-second time limits, both pipelines use static tensor shape pre-allocation. This prevents dynamic memory fragmentation and keeps resource consumption predictable throughout the evaluation.