Bohea — Engineering NotesShort engineering essays on building at the seam of software and hardware. Proof before promises.https://boheastill.com/2026-08-08T04:01:07+08:00Boheahttps://boheastill.comAI extrapolates fluency. I extrapolate emotion.https://boheastill.com/notes/fluency-and-emotion-the-correcting-pair2026-08-08T04:01:07+08:002026-08-08T04:01:07+08:00The popular split — AI does the technical work, the human handles communication — is backwards. AI and people fail in opposite directions, and quality comes from the correcting pair, not from either side. Here's the protocol I run.You can't see a refactor. So try to break the machine.https://boheastill.com/notes/try-to-break-the-machine2026-08-06T14:01:36+08:002026-08-06T14:01:36+08:00You paid for a refactor and nothing looks different — same paths, same screens. My answer: don't review the code. Run a break-it checklist against the machine — pull a sensor, feed an illegal part, watch it stop safely. How I make invisible work perceivable, even if you're not technical.I use AI heavily. The machine still can't hurt you.https://boheastill.com/notes/ai-on-machines-that-can-hurt-people2026-08-03T22:44:18+08:002026-08-03T22:44:18+08:00Every time I ask AI to review its own code, it finds new bugs. Not using it means being generationally outpaced. On software that moves a physical machine, how do you weigh that? Gates, like a type system.One email per milestonehttps://boheastill.com/notes/one-email-per-milestone2026-08-03T22:44:18+08:002026-08-03T22:44:18+08:00You want "an HMI," and that's the whole spec. I split the work into milestones so it can't spiral — and run each one with a single email. One big risk becomes small stages you just watch complete.Where Value Goes When Building Is Freehttps://boheastill.com/notes/where-value-goes-when-building-is-free2026-07-16T01:23:40+08:002026-07-16T01:23:40+08:00When a cost goes to zero, value doesn't vanish — it moves to whatever became the new bottleneck. As AI makes building cheap, value migrates to judgment, accountability, and verification. And a machine that can hurt someone still needs a person who can be held responsible.Why I Lead With a Runnable Demohttps://boheastill.com/notes/why-i-lead-with-a-runnable-demo2026-07-16T01:15:21+08:002026-07-16T01:15:21+08:00A quote is cheap and a claim is easy. Handing you something that runs — on mock data, before you spend a credential — moves the risk off you and onto me, where it belongs.Industrial Reliability: Change the Denominatorhttps://boheastill.com/notes/industrial-reliability-change-the-denominator2026-07-15T14:29:14+08:002026-07-15T14:29:14+08:00You can't count the bugs that will hurt you — they're the ones you couldn't imagine. So stop counting bugs and count consequence classes instead. How to make a machine safe when you already know your software has bugs you haven't found.A Private AI Assistant That Never Leaks: Designing the Outbound Boundaryhttps://boheastill.com/notes/private-assistant-outbound-boundary2026-07-15T14:19:35+08:002026-07-15T14:19:35+08:00'Your data never leaves your machine' is only real if you can point at the single place it does leave — and prove nothing else can. Designing the outbound boundary for a self-hosted AI assistant.Governing AI Agents That Touch Moneyhttps://boheastill.com/notes/governing-agents-that-touch-money2026-07-15T14:16:15+08:002026-07-15T14:16:15+08:00The agent proposes, it never executes. Every write with a material consequence leaves the model's reach and passes an authorization layer the agent cannot author or bypass — with an escrow flow as the worked example.