Quadravox

Quadravox is an AI video-to-style captioning agent built for the AMD Developer Hackathon (Track 2). It watches a video clip and generates precise, on-tone captions in four styles — Formal, Sarcastic, Humorous (Tech), and Humorous (Non-Tech) — generalizing across nature, urban, animals, people, sports, food, and tech content. Architecture (high level): Preprocessing — Evenly-spaced keyframes are extracted across the full clip (including its final frames) at a resolution high enough to catch fine details like signage or device type. Audio is pulled in parallel and transcribed for spoken-word context (optional, non-fatal — captioning still works if this step fails). Structured Vision Captioning — Keyframes and transcript are sent to a vision-capable LLM in a single structured call, with a fallback model on standby. JSON mode plus a strict, few-shot prompt returns all four styles as clean JSON, avoiding the chain-of-thought leakage and truncation that break naive prompting. A defensive parser and placeholder guard reject any stray or template output. Self-Evaluation & Repair — Each caption is scored on accuracy and style-match, mirroring the judge's rubric; only weak styles get regenerated, preserving the strong ones. Key features: Specific & factual — all four captions derive from the same high-detail keyframes and are prompted to name concrete specifics, so no style contradicts another. Parallel & fast — clips are processed concurrently, so the full set finishes comfortably within the time limit. Resilient fail-safes — checkpointed, pre-seeded output means results are always valid, complete JSON — no style scores zero by omission. Lightweight — a minimal containerized service reads a task file and writes results, no heavyweight dependencies.