apna_captioning_agent_v5

This agent generates styled captions for video clips across four tones — formal, sarcastic, humorous-tech, and humorous-non-tech. It runs a two-stage pipeline behind a pluggable provider architecture: first it downloads the clip and samples evenly-spaced keyframes with OpenCV, then a vision-language model (via Fireworks AI API) produces a faithful, neutral scene description. A second model ) rewrites that description into each requested style, guided by tone-specific prompts. The design cleanly separates the vision and styling stages, so any model — local or hosted — can be swapped in through environment variables without touching pipeline code. Engineered for reliability under the evaluation constraints: it guarantees a caption for every requested style, writes strictly valid JSON atomically, retries API calls with exponential backoff on rate limits, caps frame resolution to bound latency, and completes the full clip set well within the ten-minute runtime limit. The container reads tasks from /input/tasks.json and writes /output/results.json, running on linux/amd64 with no GPU required at inference time.