OmniCaption AI

Built a next-generation video captioning system using Fireworks AI that accurately understands video context, generates four distinct caption styles, prevents hallucinations, and runs efficiently on AMD hardware. To solve this, I built OmniCaption AI using a unique 13-agent architecture. Instead of asking one AI to guess, specialized agents handle strict, divided responsibilities. Whisper (accelerated by AMD ROCm) transcribes audio, while Vision models (via Fireworks AI) extract strictly verifiable visual facts from frames. A Story Builder Agent combines these inputs, passing them to specialized creative agents to generate 4 distinct caption styles, social posts, and rich accessibility descriptions. Finally, a Verification Agent actively fact-checks all outputs against the original data. Ultimately, OmniCaption AI transforms any video into a structured intelligence hub. From a single upload, it delivers hallucination-free outputs including Formal, Sarcastic, Tech Humor, and Funny captions, alongside emotion timelines and rich accessibility narratives.