nami - video caption

nami is a grounded video understanding system built to generate accurate, natural, and stylistically diverse captions from video content. Rather than relying on a single prompt, the pipeline separates the problem into distinct stages: intelligent frame sampling, objective scene understanding, temporal event consolidation, factual grounding, and style-aware caption synthesis. The system first extracts representative visual information from the video and constructs an objective timeline of observable events. This intermediate representation serves as the single source of truth for every generated caption, reducing hallucinations and ensuring stylistic variations never drift away from what is actually present in the video. Using this grounded representation, nami produces four distinct caption styles: Formal, Sarcastic, Humorous (Tech), and Humorous (Non-Tech), each maintaining factual consistency while adopting its own unique tone and writing style. The project emphasizes reliability as much as language quality. It incorporates structured validation, deterministic processing, robust error handling, and optimized inference to deliver consistent outputs across diverse video domains, including sports, wildlife, urban environments, office scenes, and time-lapse footage. Built and optimized for the AMD Developer Hackathon, nami demonstrates how a modular, grounded AI pipeline can transform raw video into expressive, trustworthy language without sacrificing factual accuracy or stylistic creativity.