video-captioning-agent

Ever wish your video captions had a personality? This project watches a video exactly once - no re-analyzing, no repeated model calls - and builds a strict, factual understanding of what's actually happening: who's there, what they're doing, how the scene changes over time. We call this the Canonical Video Report, and it's the single source of truth for everything that comes next. From that one factual report, a second model takes over and rewrites it into whichever caption styles you ask for: dry and Formal, dead-pan Sarcastic, pun-heavy Humorous-Tech, or just plain Humorous. The trick is that the writer never sees the video itself - only the facts already extracted - so the humor can get creative without ever making something up that didn't happen. Under the hood, it's two specialized models doing two different jobs: one that's built to look closely at frames and stay boring on purpose, and one that's built to have a little fun with language. We run them on dedicated, purpose-picked hardware, track every prompt and parameter tweak we make along the way, and the whole thing is built to process a batch of videos start to finish in minutes, not hours - reliably enough that it never crashes, even when a video or a request goes sideways.