Calisk-OCR and Arsa-RAG

Calisk-OCR: High-Performance Vision-Language OCR Pipeline for AMD ROCM Calisk-OCR is a production-grade, hardware-aware Optical Character Recognition (OCR) pipeline designed for the AMD x Lablab.ai AI Challenge (Mini-Challenge 2).Engineered to operate under adverse real-world conditions—such as severe glare, motion blur, low lighting, and angled perspectives—Calisk-OCR extracts precise text from complex roadway imagery, license plates, and traffic signs. Core Architecture & Key Features Vision-Language Model Engine:Powered by Meta's `Llama-3.2-11B-Vision-Instruct` and served via the `vLLM' inference engine, optimized specifically for AMD ROCm GPU acceleration. Multi-Pass Resiliency: Integrates an OpenCV CLAHE (Contrast Limited Adaptive Histogram Equalization) preprocessing layer to sharpen degraded inputs, backed by a secondary EasyOCR fallback pass to guarantee execution reliability. Strict Memory & VRAM Management:Enforces explicit GPU memory bounds (`gpu_memory_utilization=0.65`) and eager execution modes, maintaining peak VRAM usage safely below the 48 GiB limit without memory leaks or OOM crashes. Automated Standardization: Features automated text normalization (UPPERCASE conversion, noise reduction, and banner stripping) combined with a dynamic confidence scoring algorithm. Containerized Deployment: Fully Dockerized using AMD's official ROCm PyTorch base image (`rocm/pytorch:rocm10.0_ubuntu26.04_py3.14_pytorch_release_2.13.0`), delivering a self-contained CLI interface that accepts PNG, JPEG, and TIFF formats while outputting structured, evaluation-ready JSON artifacts. Arsa-RAG: AMD-powered multimodal RAG project is a multimodal retrieval-augmented generation (RAG) project built for AMD MI300X/ROCm environments. It indexes a local corpus, retrieves relevant information, and answers user questions with citations and confidence scores for GPU-accelerated, document-based Q&A workflows.