This project introduces a highly optimized, asynchronous video captioning pipeline engineered for fast hackathon evaluations. By bypassing traditional, time-consuming file downloads, the system leverages raw FFmpeg network input-seeking to isolate and extract keyframes instantly using HTTP range requests. These downscaled frames are converted to base64 and dispatched to the Fireworks AI gateway. First, the Qwen-3.7-Plus vision model generates a foundational visual description, which the DeepSeek-V4-Pro text model dynamically interprets into multiple distinct caption styles. Hardened with exponential backoff retries, the architecture cleanly eliminates network timeouts.
Category tags: