SceneSense: Adaptive Video Intelligence Agent

SceneSense is an advanced multi-modal agent designed to extract high-fidelity insights from video content. By orchestrating a highly concurrent pipeline of Vision-Language Models (VLMs), audio transcription, and stylistic LLM synthesis, SceneSense is capable of "watching" a video and describing it with exceptional accuracy in multiple tones and styles. At its core, SceneSense Abandons the standard single-frame extraction method. Instead, it features an intelligent multi-frame sampling engine that pulls multiple high-resolution frames (at the 25%, 50%, and 75% marks) from any scene longer than two seconds. These frames are analyzed using a strict 1280px resolution threshold, completely eliminating OCR hallucinations and allowing the agent to read on-screen text with perfect clarity. The visual and audio data is then aggregated into a deduplicated, chronological timeline builder. Finally, the synthesis engine routes the timeline to DeepSeek V4 Flash using dynamic temperature constraints. It enforces a strict temperature=0.1 for hyper-accurate Formal descriptions, while dynamically unlocking temperature=0.7 for Sarcastic and Humorous outputs, ensuring optimal stylistic variance without sacrificing factual integrity.