.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Draft-1 Video captioning: 404 Found
This Track 2 submission is a video captioning agent built around a two-stage pipeline designed to separately optimize for accuracy and style match, the two axes the judge scores on. Given a video clip, the agent first downloads it, samples a handful of representative frames with FFmpeg, and sends them to a vision-capable model on Fireworks AI, which produces one neutral, factual description of the scene. That description is then rewritten into four distinct caption styles—formal, sarcastic, humorous with technology references, and humorous with no technical jargon—through a single structured call to a second Fireworks-hosted model, returned as one JSON object. The pipeline includes defensive handling throughout: automatic retries and a fallback model if a primary call fails; JSON parsing that tolerates markdown fences or stray text around the response; truncation repair if a generation is cut off mid-sentence; and a guarantee that every requested style always receives a non-empty, complete caption—even in worst-case failure scenarios—so no single dropped field can zero out an otherwise-working clip. The entire agent runs as a self-contained Docker image with no environment variables required at runtime, reading tasks from /input/tasks.json and writing results to /output/results.json, matching the evaluation harness exactly. It has been tested end-to-end against the provided example clips and verified through an automated CI pipeline that builds, runs, and validates the container the same way the judge will before every publish.
13 Jul 2026
