VeloCap: High-Speed Video Agent

Our Video Captioning Agent for Track 2 bridges high-speed visual processing with context-aware language generation, turning any video into accurate, style-specific captions — formal, sarcastic, humorous_tech, and humorous_non_tech — as clean, structured JSON. System Architecture & Workflow Built in Python and packaged in a lightweight linux/amd64 Docker container, the agent runs a streamlined pipeline: it reads incoming tasks, retrieves the source video, and uses headless OpenCV to extract 24 evenly spaced frames, each resized to 640×360 for fast, token-efficient inference without losing semantic detail. A ThreadPoolExecutor processes multiple videos concurrently, cutting total runtime significantly, while retry logic with backoff covers both downloads and API calls — so a single transient failure never crashes the run. The Engine: MiniMax M3 via Fireworks AI The submission form only offers a generic "AI/ML API" category, but the model actually powering the agent is MiniMax M3, served through Fireworks AI — chosen deliberately for three reasons: 1)Native multimodality — trained from the ground up on text, image, and video together, enabling deeper semantic fusion across frame sequences than a bolted-on image-to-text pipeline. 2)Efficient MoE architecture — ~428B total parameters with only ~23B active per token, delivering frontier-level reasoning at speeds a dense model of similar size can't match. 3)Massive context window — 512K tokens comfortably holds all 24 base64-encoded frames in a single call, with no truncation or memory loss. Performance & Reliability Frame optimization and parallel execution let the agent process multiple clips and generate a full set of style-specific captions for each in roughly 16–20 seconds end-to-end. Retry logic and failsafes at every stage — download, extraction, API call — ensure the agent never crashes and always returns valid JSON, in line with Track 2's evaluation requirements.