Design the fallback before the outage

Every AI product has a dependency it doesn't control: the LLM provider, the model name, the GPU it borrowed. Across the five production apps I shipped this year, the same rule held — design the fallback path before launch, not after the first outage. RAG.NextUpgrad falls back Groq → Anthropic → OpenAI behind one port, with identical prompts. DocuLens keeps every model name in an environment variable so a retired model is a config change, not a code change. FaceVision runs WebGPU-first with a WASM fallback in the browser, and Redis with an in-memory fallback on the server. The OCR & Speech Workspace uses two speech models — one for latency while you talk, one for accuracy when you stop. FineTune Studio serves inference from a tunnelled vLLM on a borrowed Kaggle T4, so the fallback for a dead tunnel is designed in. And in every campus workshop I run, the best fifteen minutes are the same: kill the API key and watch what still works. I also build tools.scult.in — 15 free tools, 1,200+ prompts and a 50,000-skill library, free for anyone, no signup.