FourFaced

FourFaced is a video captioning agent built for Track 2 of the AMD Developer Hackathon: ACT II. Given a video clip, it generates captions in four required styles — formal, sarcastic, humorous-tech, and humorous-non-tech — using Gemma 4 for the entire pipeline, not just the final write-up. Gemma looks at the clip and grounds a factual description of what's actually happening, then writes all four styled captions from that description, reviewing its own draft against the facts when time allows. Every caption ships alongside the real grounding facts behind it, so accuracy isn't just claimed, it's checkable. Reliability was engineered under real conditions: during testing, a live outage on Google's own Gemini API led us to build a same-provider failover (Gemma 4 31B to 26B) covering every call in the pipeline, verified against a real, confirmed outage rather than a hypothetical one. The system also shows, live, which model actually answered each request. Beyond the core pipeline, FourFaced includes batch processing for up to four clips at once, a built-in plain-language guide for non-technical users, per-style selection, light and dark mode, and a companion mobile app. We also tested past the three official examples, and documented — rather than hid — a real limitation we found on fast-action content.