Wingman · in your ear and on the code
Do your paperwork in your ear
Wingman reads the technician’s and the driver’s checklists into a Bluetooth earpiece, one step at a time, and captures what they say back, so both hands stay on the work and the phone stays in the pocket. Underneath it is the same run as the app, so the record is identical whether it was spoken or typed, and the job is closed out before the tech gets home instead of at the kitchen table tonight.
Wingman also knows the code. Ask it whether 315° is over the limit for a T&P valve, or whether 2/32 of tread is too worn on a tire, and the answer comes from the code for wherever the job is, with its work shown and the section cited. That is the code itself, not a chatbot’s best guess, and it works in a basement with no signal at all.
Both halves come from the same rethink of how a field service company scales to meet what customers demand today. Wingman takes the typing and the looking up off every tech and driver, so the crew you already have can take on more, and the work grows faster than the overhead.
Say “hey wingman.” Talk. Say “over.”
Wake it with a word, answer the step it reads to you, and send with “over.” That is the whole interface. The phone never comes out.
- “hey wingman” starts it hands-free. Holding or tapping the button does the same thing when a site is loud.
- “over” sends what you have said. “faster” and “slower” tune the voice to the person, because the speed one tech likes makes another reach for the phone.
- “my jobs,” “todo,” “menu” — the same commands, always, on every job. There is nothing to learn twice.
- “what’s owed?” reads back everything that still has to be cleared before this job can close. That one sentence is the whole deferred-capture model, spoken.
- And you can just type. There is a text box right above the talk button. Voice is the fast lane, not a toll gate — a tech who hates talking to software is not locked out of the job.
Drivers get the same help with their day. Start of day, on my way, the pickup, the bank drop and end of day are read to them and answered out loud, so nobody is typing at a red light.
It rides any Bluetooth earpiece. There is no proprietary device to buy, nothing extra to charge, and nothing to lose off the back of a truck.
The run records which language each answer was given in — a technician can switch mid-job, and the answer keeps the tag. The screens themselves ship in English.
The real client, sitting idle. “What’s owed?” is a button in the top bar as well as a thing you can say.
One sentence, a dozen fields
The tech says “compressor’s shot on RTU 3, suction’s at 58, head pressure 310, unit’s iced up again, I’ve got the model and serial” — and a dozen steps fill at once, each through the exact validation a typed answer would get. No form. No order to learn. No hunting for the field somebody put on page four.
Twenty minutes of paperwork becomes two, and it happens on the drive to the next call rather than after dinner.
That distinction is the entire design, and it is the opposite of what everybody else is selling. Wingman does not decide anything. It does not invent a reading, it does not sign for the customer, and it does not get a vote on whether the job is finished. Every value it captures is one a person said out loud, checked by the same rule that would have checked it if they had thumbed it in. The validation lives on the step, not on the surface — which is why the record cannot come out weaker just because it was spoken.
Wingman knows the code
Ask the code, not a chatbot
A chatbot answers a code question the way it answers everything else, fluently, confidently and sometimes wrong, and on a job site a wrong answer means a failed inspection and a second trip. Wingman is built the other way around. Its AI has one job, which is to read the question the way a tradesperson actually asks it, filler words and all, and write it down as a short formula. Everything that has to be exact is done by our engine, from the code itself.
It is the same rule the earpiece follows. Wingman is a very fast reader and a very fast typist, and the final word always belongs to something exact: the step’s own checks when it is filling in a job, and the code itself when somebody asks it a question. Nobody else in field service has put AI to work at the level of the code.
Ask it “how many hangers for 60 ft of 3/4 pex” and
Wingman writes round up(60 ft ÷ hanger spacing for 3/4 pex) + 1,
which the engine then works out from the code. It speaks in formulas
rather than picking from a fixed menu, so it can put them together the way
a person works a problem: for each thing in this set, do this, then
average the set.
-
You ask it your way
Shop talk, shorthand and filler words are all fine. A pop-off is a relief valve, and Wingman knows it.
-
Wingman writes the formula
One short formula in a small language of our own. Reading the question is the only thing the AI does.
-
The engine finds your code
The code, then your state’s amendments, then your county’s, taken in the order an inspector reads them.
-
You get the answer, worked out
The math is done, the work is shown and the section is cited. If the reading doesn’t check out, Wingman asks instead of guessing.
Here is how that looks on real questions, word for word, including one that isn’t a code question at all and is turned down rather than answered.
-
“is 315° over the limit for a T&P”
→ check 315 <= max temperature(relief_valve)
Not compliant. You have 315 °F; the relief valve maximum temperature is 210 °F.
-
“um: is 7/16 inch the minimum supply line on a standard bathtub”
→ check 7/16 >= min supply_size(standard_bathtub)
Not compliant. You have 7/16 inch; the minimum supply pipe size is 1/2 inch.
-
“remind me, which kind for a pop-off”
→ lookup type(relief_valve)
An approved, self-closing (levered) pressure and temperature relief valve … conforming to ANSI Z21.22.
-
“How much drop over 86 feet at 1/8 per foot”
→ compute 1/8 in/ft * 86 ft
10 3/4 inches of fall.
-
“quick question, dropping 10 inch in 7 feet, what grade is that”
→ compute 10 in / 7 ft
About 1 7/16 inch per foot.
-
“hey – what size main at 1/4 inch per foot for 311 fixture units”
→ size building_drain[slope 1/4] for 311 dfu
5 inches (carries up to 480 DFU).
-
“12 service sinks and 1 vanity sink – how many dfu all together”
→ compute 12 * drain_load(service_sinks) + 1 * drain_load(vanity_sink)
12 × 2 DFU (service sink) + 1 × 1 DFU (lavatory) = 25 DFU.
-
“question for you – only got 2/32 on the tire, too worn?”
→ check 2/32 >= min tread_depth(tire)
Compliant. You have 2/32 inch; the minimum tread depth is 2/32 inch.
-
“who won the baseball game tonight”
→ none
Declined: not a code question.
The question as it was asked, the formula Wingman wrote, and the answer the engine worked out.
Your experts teach it with a spreadsheet
Every formula Wingman knows lives in a workbook, on what we call a formula card: one formula, and beside it the ways people ask for it. That means the person who teaches Wingman a trade is somebody who knows the trade. A master plumber can fill in a spreadsheet and hand it back, and that spreadsheet becomes training data, with no programmer in the middle.
Built from scratch, small enough for a phone
The model is our own, about 2 MB on the phone, and it needs no internet at all, so it works in a basement, a crawlspace or a hot attic just as well as it does in the truck.
It learned from the code, not from another AI
It was taught from two things only: the code tables, and the way tradespeople actually ask their questions. It never learned from another AI’s answers, so what it knows comes from the source.
Plumbing first, then every trade
It starts with the plumbing code and vehicle checks such as tire tread, and every other trade comes in the same way, one workbook at a time, from the people who know that trade best.
That is what makes this an AI a trade company can actually trust. The know-how lives in a spreadsheet your best people can read and correct, the answers come from the code, and every tech carries all of it in a pocket, so a question that used to mean a call back to the shop is answered on the spot.
Yes — you can defer the photos
That sentence makes software people wince. Ask anybody who has spent a day on a roof and it is the only honest way to build this. You cannot photograph a nameplate with a gauge set in one hand and a meter in the other. You cannot stop mid-braze to satisfy step seven.
So the job does not stop. It remembers.
-
The step is answered, the photo is not
The tech gives the reading and moves on. What is owed — the picture of the blower motor, the serial plate, a note somebody has to write properly — is recorded as a debt against that exact step.
-
It can be heard at any point
“What’s owed?” reads the list back. Nothing is hiding in a red asterisk somebody has to scroll to find.
-
Pay it at the tailgate, in the truck, on the drive
Or from the web review at a desk afterwards. The debt knows it was paid on another surface and clears itself — it does not sit there demanding a photo that already exists.
-
The job will not close while one is standing
Not a warning. Not a nag the tech learns to tap past. The completion gate reads the ledger, so the gate is on the data, never on the sequence — take any route through the day you like, the office gets the same complete record.
A system that quietly leaves the placeholder in is a system that manufactures evidence. Do that once and none of the rest of the record is worth defending — which is exactly the moment you needed it.
Signatures always pop a real screen
And that is mechanical rather than a promise. Every messaging channel in the engine carries a capability profile, and every one of them — RCS, WhatsApp, Apple Messages, plain SMS — says it cannot capture a signature. So the run flips to a screen instead of asking somebody to say their name out loud and calling it a signature.
Nothing closes half-done
The gate sits on the run itself rather than on a screen, so it holds on every door a job can be completed through. The app, the browser, the text thread and the earpiece all pass through that same gate, and a gate with one unguarded entrance is not a gate.
The order was yours, not ours
It is hailing, so the roof waits. The tenant is on a call, so the panel waits. Start inside and finish outside when the storm passes. Nothing was skipped — the sequence was just the one the day actually had.
The gate never takes anyone’s word for it
The completion gate reads the ledger of what is still owed, so no amount of talking, tapping or texting can close a job while a capture is outstanding. That holds for Wingman exactly as it holds for the tech.
Four ways to run the same job
A tech on a ladder wants the earpiece. A sub with no app wants a text thread. A QA lead wants a tablet. The office wants a browser. It is one authored job, and all four produce the same structured record — so you never have to choose between what the field will tolerate and what the office needs to bill.
The mobile app
Built for your own crews, and in use every working day at a plumbing shop with forty-plus techs. Full offline, because a basement or a crawlspace does not care about your roadmap. What the tech gets →
Tablet and browser
The same run on a full screen, for the office, for QA, and for the long forms and signatures that genuinely deserve a screen. This is also where a deferred capture gets paid off at a desk.
The earpiece
The technician’s or the driver’s checklist is read to them and answered out loud, with hands on the equipment and the phone in the pocket, so the last call is closed out on the drive to the next one.
A text thread
The job runs as a conversation in the messages app already on their phone. Nothing to install, nothing to log into, no seat to buy — which is the only reason a one-day sub ever follows your process. Subcontractors, over a text →
Where we would like this to go
Because a step declares what it needs rather than what screen to draw, an entire run can be heard and answered aloud — the to-do list, the steps, the captures, the finish. That is the mechanism that would let a blind or low-vision technician run the same job, on the same record, with no separate app and no lesser version of it.
We will not claim it until a technician who cannot see the screen has run a whole job with it, start to finish. It is why the engine was built the way it was, and it is where we are taking it.
The training is spoken too — eight chapters a tech can listen to, plus two cheat sheets and a playbook that reads offline.
Everything Wingman does, in one place
We sell Wingman in person, to people who ask to see things, so here it is side by side: what it does in your ear, and what it does with the code.
In your ear
- Reads the technician’s and the driver’s checklists into any Bluetooth earpiece and captures the answers, with “hey wingman” to start and “over” to send.
- Fills a dozen steps from one sentence, each checked exactly as a typed answer would be.
- Answers “what’s owed?” with every capture still standing, tracked against the step it belongs to.
- Keeps the job open until its record is complete, on every door a job can be closed through.
- Holds a required photo’s place until the real one lands, and asks once about an optional one before deleting its placeholder.
- Pops a real screen for every signature.
- Records which language each answer was given in.
On the code
- Reads the question the way it was asked and writes it as a short formula.
- Looks up the value for wherever the job is: the code, then the state’s amendments, then the county’s.
- Does the math, shows its work and cites the section.
- Asks the tech when a reading doesn’t check out, and declines anything that isn’t a code question.
- Runs on the phone with no internet at all, on a model of our own about 2 MB in size.
- Learned only from the code tables and the way tradespeople ask, never from another AI’s answers.
- Takes each new trade from a spreadsheet an expert fills in, starting with plumbing and vehicle checks.
Both halves follow one rule. The AI does the typing and the reading, and every decision rests on something exact: the step’s own checks, the ledger of what is still owed, or the code itself. That is how the crew you already have gets faster on every job without anybody having to trust a guess.