Skip to content
Vest Practices
Smart glasses, folded, seen from above

Labor in America is slowed by a lack of tradespeople. A chatbot can’t replace an electrician or a data center technician.

The trade gets passed down one apprentice at a time, and there aren’t enough journeymen left to train them. So we built one.

AGC of America workforce survey · 1,342 firms · 2025

Gloved hands working on a connection, seen from the wearer's view
01 — How it works

It sees what you see.

Every other tool talks about your work in the third person — a manual, a video, a chat window you look away to read. This sits on your face. It watches the bench from your eyes, hears the question you ask with your hands full, and answers out loud while you keep working.

Equipment racks dense with labelled modules and part numbers
02 — At every step

It tells you what to do next.

“Which of these is the neutral?” Hold the panel up and it reads the labels with you, pulls the spec for a part number no search engine has indexed, and walks the circuit back with you until the fault is found.

Arduino UNO, breadboard and probe on a bench
03 — Over time

It remembers how you learn.

It knows you landed on COM the first three times, and that you have not missed it since. It stops repeating what you own and slows down where you struggle — the thing a good journeyman does without being asked, and a manual never will.

The procedure

It catches the mistake while it is still a hand in the air.

A journeyman writes the task once, in plain English. That is where the expertise lives — which matters, because the panel in front of an apprentice is usually proprietary hardware with no public documentation to train on. The glasses do not have to recognise the world. They only have to check what the procedure asks about.

Commission a motor controlred +5V · black GND · yellow control · green load
  1. 01

    Isolate the feed.

    Looking for: Power strip plug is out of the outlet.

    sensor: ambientdone
  2. 02

    Show me a red wire. Land it on COIL+.

    Looking for: Red wire inserted at the COIL+ label.

    Saw: Landed on COM

    That's the contact, not the coil.

    caught it
  3. 03

    Yellow from COIL− to the switch.

    Looking for: Yellow wire between COIL− and the switch terminal.

    up next
  4. 04

    Red from the contact to +5V.

    Looking for: Red wire at the NO contact and the +5V rail.

    up next
  5. 05

    Green from the contact to the fan.

    Looking for: Green wire between the contact and the fan lead.

    up next
  6. 06

    Black to ground.

    Looking for: Black wire on the ground rail.

    up next
  7. 07

    Re-energize. Stand clear. Press the button.

    Looking for: Fan is spinning.

    sensor: liveup next
How it decides
01

The model suggests. The engine decides.

A language model never advances a step on its own. A deterministic state machine does, and only on evidence it can point to.

02

One good frame is not enough.

Before anything near a conductor, it wants N consecutive frames above threshold and two independent model calls that agree.

03

“I can’t tell” is an answer.

When the view is blocked it says so and waits, rather than guessing. A tool that guesses under schedule pressure gets ignored the first time it is wrong.

04

Two senses agree before you go near a conductor.

The camera confirms the isolation state and the field sensor reads ambient. Either one alone is not enough to move.

05

Every decision keeps the frame behind it.

Timestamp, step, frame, model output, sensor level, verdict, hash of the previous entry — so the record shows the work, not just the result.

Where it stops teaching and starts refusing

One sense can be fooled. So it waits for two.

A patient teacher everywhere else — but not here. Near a live conductor the camera can be shown a plug that only looks unplugged, so it wants a field sensor to agree before anyone reaches in. Flip either switch.

Breadboard, Arduino and field probe on the bench
Stop

“Hold on. The plug looks out, but the conductor still reads live. Something else is feeding it.”

Cameraclear

Is the plug out of the outlet?

Field sensorlive

Is the conductor dead?

What it says

Stop

The plug looks out, but the conductor still reads live. Something else is feeding it.

The sensor is a presence-of-voltage guardrail, not an absence-of-voltage test to NFPA 70E. The procedure still teaches the contact test and still requires it — and the camera verifies the meter reading when it happens.

The record

It remembers what you got wrong, and that you fixed it.

Every decision keeps the frame behind it. For the apprentice that is a competency record; for the supervisor it is evidence; for the next model it is training data.

Session 04 — commission a motor controlTrainee 07 · 31 min
  • 09:02PPE confirmed — eye protection, gloves3 frames
  • 09:04Feed isolated — camera + sensor agreeverified
  • 09:07Reached for COM — caught, correctedtaught
  • 09:08COIL+ verifiedconf 0.93
  • 09:19Reached for terminals, gloves offreminded
  • 09:31Re-energize — stand clearsensor: live
  • 09:32Fan spinning — task completesigned

Hash-chained · a deleted entry shows as a gap

For the apprentice

Two corrections, both stuck

A record of what they got wrong and fixed — the thing a logbook signature can’t show.

For the supervisor

Signed without standing there

Evidence of safe procedure on a task they didn't personally witness.

For the next model

Every step, labelled

Frame plus procedure state plus outcome — a dataset that comes free with the work.

Watch it run

The same procedure, start to finish.

One unedited run of commission a motor control — asked out loud, answered out loud, including the step it catches.

Demo — 45 seconds, unedited
The case

Why this, and why now.

Why a model that has to recognise everything fails on the first unfamiliar part, why the expertise belongs in the procedure, and what that unlocks for a trade that cannot train people fast enough.

Deck — 13 slides