ramanujammv1988/edge-veda

Implements a **supervised runtime with persistent workers** that manages thermal throttling, memory pressure, and session stability through a central scheduler and adaptive compute budgets—enabling 28+ minute sessions without crashes or reloads. Provides end-to-end AI capabilities (inference, vision, STT, TTS, image generation, embeddings, RAG) integrated via pull-based streaming and grammar-constrained generation, with structured observability through JSONL performance tracing and validation telemetry for debugging production failures.