VIDEO-AGENT TACK 2

This project is an automated Video Captioning Agent developed for Track 2 of the AMD Developer Hackathon. Packaged as a standalone linux/amd64 Docker container, the Node.js (TypeScript) application acts as a complete end-to-end pipeline for video understanding and stylistic text generation. Key Architecture & Features: - Media Processing: The agent dynamically reads video input URLs and utilizes ffmpeg to extract optimal visual frames (1 frame every 10 seconds). This keeps the payload lightweight and efficient while retaining enough visual context for accurate captioning. - AI Integration: The extracted frames are encoded and sent to a Large Vision-Language Model via the Fireworks AI API. The agent parses the multimodal context to understand the scene, setting, and subjects. Intelligent Fault Tolerance: To ensure strict compliance with the hackathon's automated evaluator rules, the agent includes custom retry and validation logic. If the AI model hallucinates or fails to return the exact requested JSON schema, the agent automatically adjusts parameters and retries the request until all four required caption styles are perfectly generated. - Automated I/O Compliance: The pipeline is architected to run seamlessly in headless environments, automatically pulling job configurations from /input/tasks.json and writing structured outputs to /output/results.json upon completion.