CTO building B2B SaaS with AI. Production failures, agent experiments, and lessons from scaling real systems
India
Based in India
- Tweets 26,339
- Following 447
- Followers 1,834
- Likes 2,978
I would add shipping something small to real users within the first month, since that teaches more than another tutorial.
1
Support threads double as marketing research, since each one shows the exact words users use for the problem.
1
1
9
The cache also warms with use, because the model has seen a framework's conventions thousands of times and needs fewer tokens to follow them.
4
The browser console would be my first stop, since a blank page is usually a missing env var or an API URL still pointing at localhost.
1
14
Compiler errors give an agent a precise signal on every build, which probably explains part of the shift.
2
40
After each interview I would jot down the questions that stumped me, which turns a rejection into a study list.
25
The new-hire comparison holds in review as well: a written spec and a careful read of the diff get better results from an agent than a vague ask.
9
A fast CI run that fails with a clear message gets used by agents the same way it does by engineers, so the platform teams' work carries over directly.
13
A clear offer keeps support cheap too. Every customer question starts to look like the last one.
1
11
Giving the model a failing test alongside the issue tends to work better than the issue alone. It gives the fix something to be checked against.
6
One strong coding agent plus a cheaper second model that only reviews its diffs. A separate reviewer catches what the author misses.
6
Writing evals is the one on that list a teammate can rerun, which makes it the easiest to verify in a hiring conversation.
1
1
14