Ship the 0→1, then make it survive.
A platform is a demo on day one and an uptime problem on day ninety. I have taken one from the first commit to about twenty customer deployments, which is a different job than the one people pitch.
Replace the API call when the API call is the bottleneck.
GPT-dependent detections moved to in-house models, and one feature's spend dropped 6×. A new segmentation model took inference from 400 ms to 26 ms and GPU load at five stations from 100% to about 30%. Same accuracy requirement, different bill.
Own the boring parts.
ISO 27001:2022 from a standing start. SOC 2 Type I readiness. RLS multi-tenant Postgres. Hardened Linux on x86, micro-x86 and ARM. Three-minute median first response. Upkeep is hard, and it is also most of the job.
Grow the people, then get out of the way.
As the team went from solo to nine across time zones, my week went from 80% code to 20%. The job became getting other people to their best work, and owning my mistakes out loud so they could own theirs.