Routing rules, retries, fallback chains — how AIGate doesn't 502

Upstream LLM providers fail. Groq throttles. Anthropic rate-limits. Pollinations occasionally times out on long prompts. We built AIGate assuming all three would fail at once.

Every call goes through:

1. Rule engine reads mode → picks an ordered model list
2. Retry wrapper tries each model up to 3 times with backoff [500, 1000, 2000] ms
3. Fallback chain walks down the list — if Llama 3.3 70B is down, the call automatically lands on 8B

The result: a 502 from AIGate means everyone in your fallback chain failed. In two weeks of testing we've seen exactly zero of those.