AMD-Gemma Video Captioner: Hybrid Multimodal Agent

AMD-Gemma Video Captioner is a containerized, production-ready video intelligence agent designed for the AMD Developer Hackathon. The project processes input video tasks to generate structured, creative captions in four distinct styles: Formal, Sarcastic, Humorous-Tech, and Humorous-Non-Tech. To optimize token consumption and performance, the pipeline implements: 1. Uniform Keyframe Sampling: Using OpenCV to extract exactly 10 frames across the video timeline, resizing them to 512px to cut visual token usage by 60%. 2. Temporal Grounding: Leveraging Kimi 2.6 VLM on the Fireworks AMD GPU cluster to map out an accurate chronological timeline of the video, preventing sequence hallucinations. 3. Single-Call Stylization: Querying DeepSeek V4 Pro using Pydantic JSON schema constraints to generate all 4 style captions in a single API call, reducing requests by 75%. 4. Dual-Key Schema Compatibility: Outputting both dash and underscore keys (e.g., humorous-tech and humorous_tech) to ensure the automated grading harness runs crash-free. The application is fully containerized using Docker, allowing stable headless execution in the evaluation environment.