Frierenclaw/heiter ? reverse-engineered prompt

Build me a FastAPI based inference server for an AI app that can handle lots of requests at once without blocking.

I want it to expose an OpenAI compatible chat completions endpoint at /v1/chat/completions, including streaming responses, so I can plug it into existing clients easily. It should also have lightweight text to speech support using Edge TTS, with low latency and no paid API dependency. Make it feel like the backend brain of a larger AI ecosystem, so it can manage concurrent sessions cleanly with asyncio and background workers.

Use Python and keep it simple to run with Docker and a .env file for configuration. I’d like it organized in a way that makes it easy to extend later, and if you need to check current docs for FastAPI, OpenAI style streaming, or Edge TTS, go ahead and look them up online. Add sensible defaults, clear error handling, and make sure it works well as a high throughput service.