NotesQuestions and learnings
Questions, answered
The questions a hiring manager or a team lead tends to ask about running AI in production, answered from the systems on this site, and the learnings that came out of building them.
What these are
Not a blog and not tutorials. Each one is a question I have been asked, or a thing I had to work out, written down once so it can be read rather than repeated.
Every answer here comes from something I built and ran. When the work teaches me otherwise, the note is rewritten rather than defended.
Questions
Short answers first; the note behind each one has the reasoning.- Can AI work on confidential material without it leaving the building?Yes, if the reading is done by a self-hosted local model on hardware you control and only paraphrase travels to a frontier model for judgement. Split the work by kind, not by size, and the cost stays flat too.self-hosting · 1 min
- Do you have a favourite LLM?No. The models have different strengths, and the skill is knowing which to reach for and when. On my own systems one model researches and drafts, another checks and audits, a third builds the machinery, and a local one does the volume.models · 1 min
- How do you know an AI agent actually did the work?Every scheduled duty names the receipt that proves it ran, a file, a database row or a verify script's output, and the morning briefing counts receipts rather than claims. A finished model turn is not evidence.agents · operations · 1 min
- How do you stop an AI agent publishing something it should not?Make approve, publish and promote human-only at the database, with a trigger that refuses the status change unless the caller is marked human. Instructions can be ignored; a trigger cannot.agents · 1 min
- How do you use AI in your own development work?Daily, inside the existing review path. Browser agents to reproduce and locate live-site problems, AI on QA, two models against each other for scoping and build plans, and AI-assisted development for in-house products only. Client code is written by the developers who quoted it, and nothing ships without a person deciding it should.development · 1 min
- What does AI implementation mean in practice?Taking AI from an experiment someone ran once to a working process that runs on a schedule, with a person at every gate, a receipt for every run, and the rules enforced in structure rather than in a prompt.implementation · 1 min
- Where do you start with a team that has only experimented with AI?With the one job that costs the most elapsed time and has a clear receipt, put inside the workflow the team already has. Decide who may approve before anything runs, watch the first run, and only then give it a schedule.implementation · 1 min
Learnings
What building the systems taught, including the parts that went wrong.- Incidents, and the rules they becameA dated list of what went wrong while running AI agents in production, and the structural rule each one turned into. Appended to, never rewritten.agents · operations · incidents · 4 min
- Rebuilding this siteWhy the old single page had to go, and what is replacing it.this site · design · 1 min
- The first AI-assisted build, and what it brokeMy first site built with a model at the keyboard was a Tailwind project that taught one lesson above all others. The model kept breaking one thing to fix another, and the cure was structure around it, not a better prompt.development · learnings · 1 min