Engineering Notes — Hardware, PCBs and Verified AI

A terminal-style panel listing four passing test names from byitl's review engine, ending in Tests 20 passed. Caption: Don't buy the demo. Buy the test.

Hiring an AI developer? Ask for the test

Eight questions to ask an AI developer before you hire one, starting with the one that sorts them: how will we both know it works?

November 12, 2026 · 3 min read

Inside the HC-01 humidor controller: the custom green circuit board, an SHT40 humidity sensor and a small fan, wired together in a birch enclosure

What it costs to prototype an electronic product

A working prototype of a new electronic device costs thousands, not hundreds. Here is what sets the price, and what byitl charges.

November 11, 2026 · 4 min read

Four boxes left to right: Rules, Model (highlighted magenta), Check, Person, joined by arrows. Caption: Rules almost everywhere. A model in one place.

AI agent vs Zapier or n8n: where the model belongs

Most of a business workflow is rules that Zapier or n8n run fine. Put a model only where the input has no fixed shape, then check its output in plain code.

November 11, 2026 · 5 min read

The assembled HC-01 humidor controller in its birch plywood case, lid engraved with the byitl cigar mark, beside its printed bead drawer full of humidity gel and the jar the gel came from

It works. It is also too big.

Part five. HC-01 is running in a real humidor: 522 logged samples, agreement with an independent hygrometer, and one closed control loop.

October 6, 2026 · 6 min read

Two rows showing the same probe output: a dark crossed-out panel above and a lit panel below, both printing the identical SENSE CS OPEN line

The probe said OPEN. The display was working.

Part four. My J5 continuity probe reported OPEN on a board whose display was provably working, and a bring-up flag shipped on the demo unit.

October 5, 2026 · 6 min read

Two rows of five board outlines: the rev D row marked three dead and two partial at 0 of 5, the rev E row all five highlighted at 5 of 5

The annular-ring fix held: five of five boards work

Part three. Same design, one inverted flag put back, a different assembler. Five of five rev E boards flashed, drove the display and read the sensor.

October 3, 2026 · 8 min read

A small dashed rectangle labelled GUESSED, about 12 by 10, sitting inside a much larger solid rectangle labelled ACTUAL 25.4 by 17.8 millimetres with four M2.5 mounting holes.

The agent guessed a sensor's size. We built a mount around a part that doesn't exist.

You cannot prompt a model out of guessing. You can make the guess fail a gate: require a provenance field and reject any number without one.

September 21, 2026 · 4 min read

Two rectangles. One solid, one dashed and empty, labelled material that was never made. Their intersection is annotated as equal to zero.

Our QA agent reported zero interference. The part had a hole in it.

An interference check measures overlap between solids. It cannot see material that was never there — so it passes a part with a wall missing.

September 21, 2026 · 4 min read

Three boxes: design, QA gate, reviewer. The path carrying the verdict and history into the reviewer is cut and marked with an X. Only the artifact path remains.

The third reviewer has to be blind

Hand an agent your hypothesis and it hands the hypothesis back, confirmed. The only review worth running is one that never saw the first two.

September 21, 2026 · 4 min read

A number line. A shaded band spans 2.85 to 3.00 with a caliper reading of 2.925 inside it. The converted one-eighth-inch figure of 3.175 sits outside the band, marked with an X.

We corrected 3.0 mm to 3.175 mm. The calipers read 2.925.

Our spec said 1/8 inch Baltic birch is 3.175 mm. Correct arithmetic on a wrong label. The stock measured 2.925, and the groove was built on the error.

September 2, 2026 · 6 min read

An ESP32-S3 board inside an enclosure wall, with the 2.4 GHz antenna keep-out zone shaded and a screw boss sitting inside it marked with an X.

ESP32-S3 enclosure design: give the antenna its 15 mm

Espressif asks for at least 15 mm of clearance around an ESP32-S3's PCB antenna inside the housing. What that means for a 3D-printed case, with sources.

August 15, 2026 · 6 min read

Cross-section of a through-hole joint. Left, pads and solder fillets on both sides of the board. Right, the solder-side pad is missing and marked with an X.

We told the fab the missing annular rings were intentional. They weren't.

Part two. The fab asked about one thing before building. We said it was deliberate. It was a parser bug, and it explains the open display headers.

August 6, 2026 · 8 min read

Cost breakdown of one 3D-printed prototype enclosure: $1,589 total, of which about 99% is design and labor (18 hours of CAD plus finishing at $75/hr) and just $2.64 is filament.

What a 3D-printed prototype enclosure actually costs in 2026

Filament is under half a percent of the bill. Here's the itemized total for a palm-sized ESP32 enclosure, design time included.

July 31, 2026 · 10 min read

Break-even chart: 3D printing has no tooling cost but never gets cheaper per unit, while injection molding starts at a $10k tool then costs a few dollars per part — the two cross at roughly 500 to 1,000 units a year.

From 3D-printed enclosure to injection molding: when to switch

A single-cavity prototype steel tool for a palm-sized enclosure runs about $10k — we've had our own parts quoted. Here's the break-even math on when it pays.

July 27, 2026 · 10 min read