Rambler

Rambler is an AI-powered video captioning agent for short-form clips (30s–2 min) that goes beyond basic transcription by performing multimodal analysis of visuals and audio. It samples frames at 1 fps, extracts mono 16 kHz audio waveforms, and sends both to gemma-4-31b-it via Fireworks AI. This keeps the system simple while preserving strong video understanding. Rambler follows a human-editor style workflow with three stages. First, the model performs perception, generating a detailed scene understanding that captures on-screen visuals, facial expressions, environment cues, vocal qualities, and emotional tone. Second, it turns that analysis into four caption styles: formal broadcast narration, sarcastic commentary, humorous tech with programmer references, and humorous non-tech with everyday jokes and pop-culture hooks. Third, an evaluation step scores each caption for accuracy and style fidelity and automatically regenerates any output that falls below a quality threshold. The project includes an interactive Streamlit web app with drag-and-drop video upload, real-time pipeline progress, dual caption views (overlay or comparison cards), and confidence radar charts for accuracy and style-match scores. Rambler also supports a Dockerized headless mode for automated workflows, batch processing, and hackathon evaluation, producing structured outputs such as JSON and subtitles without the UI. Rambler is built for content creators, accessibility teams, social media managers, and video producers who need fast, tone-specific captions from raw video. Its modular stack includes gemma-4-31b-it, Fireworks AI, Python, Streamlit, OpenCV, FFmpeg, and Docker, making it easy to run locally or in the cloud.