I build small SaaS products, and the machinery that decides which one gets worked on next.
Mostly I'm not hand-coding any more. I design the architecture and the guardrails, then unattended agent sessions do the mechanical work, and nothing counts as done until it's verified against real evidence. That green wall below is almost all private repos, and most of it is agent-authored. The volume isn't the interesting part. What's interesting is what the guardrails refuse.
-
Agents work here unattended, and when something needs my judgment they DM me with the options, so I either tap one or type a direction they didn't offer. We build it together from there, and once it ships it runs its own loop: production errors get filed, fixed and deployed without me in the middle of it, and I come back in only where there's a call to make. Destroying data is the one thing they never do alone.
-
Coverage is a smoke detector, and one repo-wide number lies in both directions. One product here reads 66.91% overall, which tells you nothing: the domain is gated at 99 and reaches 99.22, the adapters sit at 0 with the reason written beside the exclusion, and each gate is set a little under what the suite actually reaches so it fires on a regression instead of on noise. I started doing it this way after finding 300 passing tests, an architecture guard, and the two rules the product exists for at 13%.
-
Reading a whole file to find one thing is the expensive habit, so I built a grep-able digest of every symbol and the line it sits on. One model file here costs about 26,000 tokens to read, its digest costs 2,900, and the answer to "what runs when this record saves" costs 140. There's no daemon and nothing to remember to re-run: it reconciles against the working tree on every call, because a copy that syncs on a timer is a second source of truth, and between the edit and the sync it answers confidently and wrongly.
-
In this loop the products harden over time on their own: an incident produces a rule, the rule gets a test, and every run after that carries it. What's left for me isn't writing the code any more, it's noticing when a hardening stopped at the product it happened in and never travelled, and then making that noticing automatic too.
TypeScript on Cloudflare edge, Rust for the rigorous core, Rails over deep PostgreSQL, Python for the LLM services. Production SaaS with the un-fun correctness in it: billing, credits, subscriptions, webhooks that fan out.