My SaaS was silently broken for four weeks because a hosted LLM model ID was retired — so I moved to local inference

Short summary

A solo SaaS founder discovered their weekly brief feature had been silently broken for four weeks because a hosted LLM model ID was retired and the API returned 404s that got swallowed. The fix: migrate the weekly brief to a local Qwen model already running on a Mac Studio M4 Max, eliminating the external API dependency entirely. Key lessons include monitoring for absence-of-output (not just errors) and treating hosted model IDs as dependencies with someone else's lifecycle.

• •Hosted LLM model ID retirement caused a silent 4-week outage with no alerts
• •Migrated weekly brief generation to local Qwen model on Mac Studio M4 Max via MLX
• •Added healthcheck that fails loudly if expected output doesn't appear by deadline

Generated with AI, which can make mistakes.