Back to Chen Wei
Evals before anything — the only ML advice I'd give a founder
July 18, 2026
If you're building an LLM product at a small team, spend the first two weeks on evals. Everything else is faster after that.
Founders ask me what infra to buy. Wrong question.
The right first step
Write 50 test cases. Ship a golden set. Now every prompt change has a scoreboard. This is worth two full-time engineers.
What NOT to do
Do not spend week one picking a vector database. That decision is reversible; a bad eval baseline is not.