I need a lightweight but highly accurate AI routine that can scan both still images or both live-streamed video & recorded videos flagging any object detection errors, noticeable colour shifts and mis-alignments in real time. The goal is simple: feed the system a stream, receive instant visual or JSON feedback on anything that doesn’t match the reference parameters I supply. Needs to be self hosted no external streaming services. My ideal workflow: • You provide a Python (or C++) module that I can drop into an existing pipeline. • It ingests common formats—JPEG/PNG for images and H.264/RTSP for video—without external conversion steps. • Object detection must identify user-defined classes with at least 95 % precision/recall on a small validation set we’ll share. • A secondary pass corrects colour variance to sRGB and highlights any frame-to-frame mis-alignments in millimetres or pixels. Acceptance criteria 1. Demo script that runs on CPU or CUDA GPU and processes a 30-second sample feed in under real-time. 2. Clear README covering dependencies, model weights and how to retrain on new classes. 3. Output examples: overlayed frames and a structured log (CSV or JSON) listing detected objects, colour shifts and alignment offsets. If you’ve tackled similar real-time QA or inspection projects, I’d love to see your approach and estimated timeline.