TL;DR: By now I focused on Esuite proving outcomes, cutting cost and closing security gaps, including some I found myself. Behind it I made a rule: if a better model makes the old design wrong, rebuild; if it only makes it slower, patch.
Era 5, October: prove it, pay less, lock the doors
Features which I would theme as: Prove it: outcome verification. Talk better. Pay less. Retire gracefully: the V1 tables moved into a separate schema rows instead of being dropped; it's reversible, because I've been wrong before. And finally (please before proceeding, let the software engineers laugh at me here) Lock the doors: my mindful internal audit found that our public anon key could read tens of thousands of rows of access logs. A migration closed it, and we found no sign anyone used it. We (me and my agents) then added a CI guard so no new table ships without RLS, and dropped a legacy users table we no longer needed. I'd rather write this paragraph and have a good laugh at me.
I can't read a stack trace like an engineer. I can read a threat like a founder whose users trust her with their bank balance, their debts and, sometimes, their family problems. So I debug in a different direction. I don't wait for things to break. I go looking for trouble - sorry, I mean the places where trust could leak.
Why I kept throwing away working code
Speed is the startup religion. I chose substance. Because small businesses don't easily forgive and they don't forget. In Nigeria, a bad experience doesn't stay a support ticket. It becomes a viral note whether you were wrong or not.
But monumentally tho, tools changed a lot between January and August. So Esuite's doc gen and decks got good enough to be worth putting a provenance system behind. By September, some T-S-T (text-to-speech-to-text) models and coding agents were like cosplaying AGI - and now Jack Dorsey's Block and its just-released Dessa, where critics are saying AI shouldn't make human gestures. Haha!
So applying the rule: if a better model makes the old design wrong, rebuild. If it only makes the old design slower, patch - it looks like: deterministic engines, one source of truth, auditable signals, lower latency, quit maintaining what isn't ours to maintain. Fewer handoffs, fewer drop-offs. Each one bought trust I can't buy back later.
Principles I'd defend in a board meeting
Deterministic where money and trust live. Usage and cost of better LLMs. Scores, stages and assignments are code, but explanations are AI. Typed capabilities, governed actions. Humans verify what matters. Measure outcomes, not activity. A finished task needs evidence. Homework over instant analysis - a consultant who answers in two seconds hasn't thought about your business. Financing goes where the value is created. Trust is a currency to bank on.
What I got wrong (the changelog remembers)
- I over-iterate on visuals before the core value is locked. Pixels are my comfort zone as a designer.
- I stack too many changes into one brief. Twelve good ideas in one prompt can still be a mess.
- The April cron incident; guards first, automation second.
- Analytics, built and removed twice.
- I don't read raw error messages myself. I read what they mean. That works until it doesn't, which is why the next senior hire is technical.
So the world has upon them a governed system with a paper trail, built by a founder who designs in systems and rebuilds when the evidence says so. It already has businesses paying to pilot it. The CTO I hire next inherits a codebase with docs, tests, a security guard in CI and a changelog that explains every decision. They inherit a system, not a rescue mission.
I still can't make that ping-pong ball bounce. But Esuite comes back every time a business owner needs it, and that's the bounce that pays charley.
Previously: Delete the First Codebase on Day Two