VoxFrame: Multimodal Narrative Synthesis Engine

VoxFrame is a state-of-the-art cognitive video analysis and narrative generation system designed for the AMD Developer Hackathon: ACT II (Track 2). Standard video captioning models often suffer from hallucinations and style drift. VoxFrame resolves this by implementing a rigorous "Comprehend-Verify-Compose-Audit" execution pattern. First, our media processing pipeline uses zero-latency FFmpeg for keyframe extraction and audio demuxing. We then use Groq's high-speed Whisper Large V3 engine to transcribe spoken audio. Instead of prompting the AI to write creative captions directly, Stage A extracts subjects, setting, and motion details. This context is verified against visual keyframes to eliminate hallucinations. Stage B then synthesizes the verified context into four distinct tones: Formal (strict third-person), Sarcastic (witty irony), Humorous Tech (using a specialized vocabulary list of developer terms), and Humorous Non-Tech. Finally, the engine audits the results. If the grade falls below a 0.65 threshold, VoxFrame initiates self-refinement, generating alternative candidates and selecting the highest-quality version. The project features a premium FastAPI-based dark-mode web dashboard for drag-and-drop interactions and is fully packaged inside a portable Docker container for zero-setup deployment.