AMD Caption Agent

The AMD Caption Agent solves the critical "single point of failure" issue in automated video content pipelines. In production environments, relying solely on cloud-based AI APIs for transcription and captioning often leads to silent failures and costly downtime. Our solution implements a robust, multi-agent orchestrator that leverages the high-speed processing of Fireworks AI (powered by AMD Instinct GPUs) for primary tasks while maintaining an intelligent, on-device fallback mechanism using "faster-whisper" and local AMD CPU compute. The pipeline architecture is designed for extreme resilience. If the primary cloud API fails, the system automatically pivots to local inference. Furthermore, for non-audio or visually complex content, the agent integrates Llama-Vision to extract keyframe context directly. The workflow utilizes a sophisticated "Critic Loop" where specialized copywriter agents (Formal, Sarcastic, Tech-Humor) draft captions, which are then evaluated and refined by a Critic agent until they meet strict quality thresholds. This ensures production-ready, high-quality output regardless of network stability or cloud availability, making it an essential tool for automated broadcasting and social media content creation.