.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Express Caption
Express Caption is a video-captioning agent built for AMD Express Track 2. Give it a short clip (30 seconds to 2 minutes) and it writes a caption in four requested styles: formal (precise and objective), sarcastic (dry and deadpan), humorous_tech (developer in-jokes), and humorous_non_tech (warm, jargon-free wit). Rather than describe the video and then restyle text, it captions in a single grounded pass: it samples frames densely across the clip (~1 per second) and, in one multimodal call, the model writes all four styles directly from what it sees, so nothing is lost in a handoff, and temporal coverage drives accuracy. A distinctness check retries if the four voices come back too similar, and every requested style is always returned. The scored deliverable is a headless Docker agent: it reads tasks from /input/tasks.json, streams frames straight from each video URL with ffmpeg (downloading only as a fallback), and writes /output/results.json with per-style fallbacks, an atomic write, and an adaptive time budget so the output is always complete and valid. Built with Fireworks AI (Kimi K2.6, a vision-language model) and ffmpeg for frame sampling.
13 Jul 2026
