CaptionMan

CaptionMan is an evidence-first video captioning system built for AMD Developer Hackathon ACT II Track 2. Its moat is the Caption Court pipeline: the system does not simply ask a model for captions and return the first answer. It extracts visual evidence from the video, generates multi-style caption candidates, checks them for grounding and quality, repairs weak outputs, and selects a final verdict-ready caption for each requested style. The project supports the official judged Docker workflow. The public linux/amd64 image reads /input/tasks.json, writes /output/results.json, validates the Track 2 schema, and uses a temporary hackathon-only Fireworks credential for runtime inference. CaptionMan is designed for hidden judging clips, with safeguards against generic captions, filename leakage, contact-sheet artifacts, frame-count language, unfinished sentences, and overly artificial phrasing. CaptionMan also includes Gemma-aware specialist routing. Gemma is configured as part of the repair/model-routing strategy, while the production route uses a champion model setup selected for reliability on the final Docker submission. This gives the system a practical balance: strong visual understanding, polished caption writing, and a transparent repair/verdict layer that makes the final output more dependable than a basic one-shot caption generator.