simple VLM

This documentation outlines the development of a **specialized AI agent** designed to generate high-quality video descriptions for the **AMD Developer Hackathon**. The system utilizes a **two-stage pipeline** that first extracts keyframes to analyze visual data before passing that information to a text model for **creative synthesis**. By separating perception from writing, the agent delivers captions in **four distinct tones**—ranging from formal to humorous—while ensuring no internal "thinking" steps appear in the final result. The architecture is **highly optimized**, featuring a lightweight container that triggers fast startups and maintains **strict schema compliance** to prevent errors during evaluation. To ensure reliability, the software includes **robust fallback mechanisms** and dynamic model selection to handle various technical environments effectively. This project prioritizes **temporal awareness**, allowing the AI to accurately track movement and sequences across short video clips.