"MADHAV AI LAB team on AMD AI League Hackathon"

MADHAV Lens is an OCR application developed for Mini Challenge 2 of the AMD AI Academy Challenge. It uses Qwen2.5-VL-3B-Instruct with PyTorch and AMD ROCm to read vehicle registration plates and road signs. The application supports PNG, JPEG, and TIFF inputs and writes extracted text in the required JSON format. A persistent GPU worker loads the model once, while a lightweight command-line client handles individual image requests. For license plates, the prompt requests the registration number without surrounding state banners or slogans. For road signs, it requests the printed text in reading order. Testing on an AMD Radeon Pro W7900D correctly transcribed all 10 public PDF samples and all 54 cases in a supplementary sample-based test suite. These results are development checks, not hidden-set evaluation scores. Additional real-world testing revealed some missed text and character errors. The Docker image has been built using the mandated ROCm base, and its library imports and size have been checked. Final GPU-container execution and uncached startup validation remain pending.