Field Notes
-
Why Firmware Is More Flaky than Software
A customer asked how confident I was that their firmware was ready to ship. I said it should be fine for the first hundred thousand units. Here's the math behind that answer — why firmware that passes every product test still acts weird in the field, and what to ask for if you're paying someone else to write it.
-
The chip wasn't slow — my benchmark was lying. Twice.
My benchmark ranked a 200 MHz dsPIC as one of the slowest microcontrollers. The chip was fine — it was still on its 8 MHz reset oscillator, so every number it produced was 25x low. Then a compiler quietly deleted 31 of every 32 additions on the 8-bit PICs. Here are real measurements from 16 microcontrollers after those issues were fixed.
-
The 16-port test rig that AI agents use more than I do
Unit-testing a pile of microcontrollers over USB fell apart at five or six boards, so I laid out a 16-port hub board that can switch and monitor the power on every port. I built it for unit tests. Now AI agents use it far more than I do.
-
Don't ask AI for a STEP file, part 2: it printed right the first time
In July I said don't ask an AI for a STEP file — ask it to write the program that makes one. I took my own advice for a real part. I kicked off the run on a Friday afternoon and walked away; ten hours later the AI had measured my PCB from the Gerbers, written the CAD program, and proved it printable. The enclosure printed right the first time.
-
Using AI for Firmware Development
A practical workflow for using AI on real firmware — get your docs into Markdown, run deep debug sessions with the best model you've got, and keep your engineers firmly in charge of the product.
-
AI for Engineering Leadership — They're Tools, Not Your Team
Once you can delegate real work to AI agents, the instinct is to manage them like engineers. That instinct is exactly backwards — and it'll wreck your schedule and your judgment if you let it.
-
AI & Software Development: Should We Be Scared?
It was November when I first got worried AI was coming for my job. Nine months and a lot of tokens later, here's where I landed — what AI actually changes in firmware work, what it doesn't, and why your job is safe.
-
Data Faults — The Bugs That Hide in Plain Sight
Your C is correct, the compiler is correct, and the product still fails in the field. That's a data fault — the interrupt fired at exactly the wrong moment. I break one down to three instructions of disassembly and show why volatile didn't save you.
-
Keeping Your Data Out of AI Training
If you're worried about your code and product data feeding someone's model, you're right to be. A practical look at what actually leaks, what the settings really mean, and how to use AI without handing over your IP.
-
Automated Firmware Testing with AI
Testing is the weak spot in a lot of embedded projects — manual, brittle, hardware-dependent. Here's where AI helps close that gap, and where it still needs a human.
-
AI Comparison with PCB Symbols and 3d Models
I had six different models make me PCB Footprints and 3D models (STEP) using nothing but the datasheet and the same prompt. These are tasks that are time-consuming and can make or break a board layout.
-
Token Usage
How I run 4.5 billion tokens a month for hundreds of dollars — and why context management, not raw spend, is the real lever.
-
Welcome to Field Notes
A quick hello, and what I'll be writing about here — connected product development, firmware, hardware, and what AI actually changes about the work of shipping.