Most AI apps are locked to one model. Chymera uses an intent classifier to route each message — coding queries go to Llama 3.3 70B, reasoning to Qwen QwQ 32B, and general/search tasks to Gemini 2.5 Flash. It remembers you across sessions via semantic memory (Mem0), streams tokens in real-time via SSE, and never goes down on rate limits thanks to automatic API key rotation across a pool of keys. Built solo. Three-tier production architecture. Actually deployed — not a demo.