KryinCaption Turbo: High-Speed Captioning Agent

KryinCaption Turbo is a state-of-the-art agentic video captioning system engineered for extreme execution speed and high descriptive precision. Built to excel under evaluation limits, the agent replaces traditional slow rendering pipelines with a lightweight, streaming-first architecture. It directly grabs 24 evenly-spaced frames from raw video streams using cv2, bypassing disk-heavy downloads. The pipeline employs a two-stage vision analysis model: Stage 1 processes all visual frames in parallel to construct a deep scene understanding (capturing location, mood, visual ironies, and narrative arcs); Stage 2 generates four distinct custom caption styles (Formal, Sarcastic, Tech Humor, Non-Tech Humor) in a single API call based on that analysis. To guarantee quality, the system runs a lightweight critic loop where a Gemma model reviews and scores caption authenticity, triggering targeted rewrites only for low-scoring styles. It features a robust multi-provider fallback chain (Fireworks AI -> Google Gemini -> OpenRouter) to handle API outages or rate limits, and uses active reasoning controls (minimal thinking budget) to keep latency low. The entire system is packaged as a lightweight, production-ready Docker container designed to easily integrate with standard batch evaluation pipelines, outputting structured JSON results rapidly and securely.