SwiftCap

SwiftCap is a video captioning agent built for Track 2 of the AMD Developer Hackathon. Given a video URL or upload, it generates four distinct captions per clip: formal, sarcastic, humorous-tech, and humorous-non-tech, all grounded in the same sampled frames so every style stays faithful to what is actually on screen. The pipeline streams video on the fly with OpenCV and yt-dlp instead of downloading full files, samples one frame per second, filters out corrupted or blank frames, and resizes for token efficiency. A single multimodal request to MiniMax M3 on Fireworks AI then returns all four styles at once, with strict JSON validation and an automatic stricter retry if parsing fails. Reliability was treated as a first-class feature: every clip is processed in isolation so one bad video cannot zero out an entire submission, and a wall-clock budget guard keeps every run safely under the ten-minute limit. The product itself is a full-stack app: FastAPI and PostgreSQL on the backend with JWT authentication, and a React and Vite frontend where users submit clips and watch captions arrive in real time.