The number itself isn't the story. The story is what makes it possible without cutting senior-level review on every line.
Three patterns that hold across all 20+ repos:
1. AI gets deep autonomy on bounded surfaces. A task with a clear "done when" — a tested module, a typed API, a documented schema — is an AI task. A task with ambiguous scope is a human task. The boundary moves over time as the agent gets better, but it always exists.
2. Review discipline is non-negotiable. Every PR runs through /security-review and /performance-review before merge. Tests must pass. Types must hold. The agent writes the code, but the human owns the merge. I have not relaxed this rule once in seven months.
3. Infra is repeatable, products are not. SST v3 templates, IAM Identity Center patterns, shared CI workflows — those get extracted. Product code stays per-repo. The reuse layer is deployment, not features.
What this isn't:
- A vibe-coding play
- A way to fire engineers
- A path to skipping fundamentals
What it is: a way for someone with senior-level fundamentals to run a one-person engineering practice the way a small team used to.
Below the surface there are about five categories of work doing the heavy lifting. I'll write each up: the BTO real-time fantasy-sports platform, the MASS Lead Connect voice receptionist, Token Holder, KnowMessenger, and the trading engines. Each one shipped because the loop above held.