Petra

CineGraph Captioner is a 3-stage video captioning pipeline built for Track 2. The user uploads an MP4 video (30s-2min) through a Streamlit interface. Stage 1 extracts the audio track using moviepy (ffmpeg). Stage 2 transcribes the audio via Fireworks Whisper-v3, producing a text transcript. Stage 3 feeds the transcript to Fireworks Llama 3.2 3B Instruct with four distinct system prompts — formal, sarcastic, humorous with tech references, and humorous without tech references — generating one caption per style. The UI displays the transcript and all four captions in expandable cards with a real-time status indicator showing pipeline progress. Containerized via Docker Compose with all Fireworks API access. Fine-tuning is supported as a future enhancement path.