EyeQ 3

EyeQ 3 is a collaborative multi-model AI system designed for automated video caption generation. The pipeline combines video understanding, frame-level validation, and style-aware caption generation to produce accurate captions in four distinct writing styles: Formal, Sarcastic, Humorous Tech, and Humorous Non-Tech. The workflow begins with Gemini 3.5 Flash analyzing the complete video to understand the scene, subject, actions, and context. Representative frames are then extracted and validated using Qwen3-VL to improve visual consistency and reduce unsupported observations. Using this validated information, Claude Sonnet generates polished captions tailored to each requested style. The project is implemented using Python, OpenCV, and Docker with a modular architecture that separates orchestration, video summarization, visual validation, prompt management, and caption generation into dedicated components. The container follows the Track 2 evaluation interface by processing video tasks from /input/tasks.json and producing structured JSON results in /output/results.json. EyeQ 3 is designed to generalize across diverse video content while producing context-aware, stylistically consistent, and evaluation-ready captions through a collaborative AI pipeline.