Wiring an LLM into a real feature instead of a demo

01

A model demo is easy: you write a good prompt, feed it a clean example, and it looks magical. A shippable feature is a different problem because the model has to work for messy, average, sometimes-adversarial input.

02

For the LinkedIn copilot, the hard part was never getting Gemini to draft a reasonable outreach message. It was deciding what the model was not allowed to do on its own: no message sends without approval, no scraping outside user intent, no silent retries that could create spam.

03

The other underrated piece is treating the model's output as untrusted input to the rest of your system. Validate the shape, cap the length, and have a fallback path for when the response does not parse the way you expected.

04

The takeaway that generalizes: the model is maybe 20% of the engineering effort on an AI feature. The other 80% is product discipline, failure handling, and clear human control.