Veedra
Veedra is a Dockerized video captioning agent built for AMD Developer Hackathon: ACT II, Track 2. Given a short clip, it produces four stylistically distinct captions: formal, sarcastic, humorous-tech, and humorous-non-tech. The pipeline is built around accuracy. Veedra samples a handful of key frames at temporal midpoints, then a vision model writes a structured scene brief. A second verification pass checks that brief against the same frames and strips out overlays, transitions, and any detail not truly on screen. Only then are the four captions written from the verified facts, each aware of the others so the outputs never sound alike. It reads /input/tasks.json and writes /output/results.json, runs on linux/amd64, and guarantees every task and style an answer. Powered by Fireworks AI using MiniMax M3 for vision and Kimi K2P6 for captions. Open source, MIT licensed.
Category tags:.png&w=96&q=75){kind=small}
For Innovators & Creators