Built Sep 20, 2026 · five days after Jev's release · real numbers from my machine

I Built a Talking AI Mirror With Jev AI

I built a talking AI mirror with Jev AI, and it picks your whole outfit in under one second.

You say where you're going, and the mirror dresses you for it and changes the room behind you.

It makes seven decisions from one sentence, and it grades its own work.

When it gets the outfit wrong, it fixes it before you ever see it.

The same trick can sort your inbox, score your leads and route your customers at the speed of a blink.

Stick with me, because one sentence I wrote made this mirror ignore me completely, and I'll show you the fix.

Say where you're going. The mirror dresses you.You say one thing"I'm giving akeynote tomorrow"Jev decides7 picks inunder a secondIt grades itselfbad grade =second tryThe mirror changesnew outfit,new roomEvery pick comes with a number that says how sure it is
The actual sources ↓
§1Watch it work first

Five sentences. Five outfits. Five rooms.

the real app on my machine · real Jev calls · recorded as it ran

What you're watching: five real requests typed into the mirror, one after another — gym, beach, drinks, a cold client meeting, a keynote. Every pick on the right is a live Jev answer.

§2Where the idea came from

Nailthy Tang built it first, with the TypeSafe team

Drape — the experiment that started this$0.0011her cost per decision~620 msher time per decision3things I added on topShe talks. Jev reads it. Her outfit changes.
The three things I addedThe room changes too8 placesIt knows when to stay stillsmall talk = no changeIt grades its own outfitand fixes a bad oneSame idea, pushed a bit further
§3What's actually new here

Jev doesn't write. It decides.

A different kind of AI modelThe situationwhat I just said+ what I'm wearingA questionwith a list ofallowed answersJev picks onein milliseconds+ how sure it isa number youcan act onReleased September 15, 2026 by TypeSafe AI
"At OpenAI, I helped build the methods that made language models useful at following instructions and talking with people. That work ended up as the research behind ChatGPT."— Diogo Almeida, founder, TypeSafe AI · launch essay · Sep 15, 2026
§4How the mirror works

An ear, a brain and a tailor

Three parts, and only one of them thinksThe earmy voicebecomes wordsThe brain · Jev7 decisionsfrom one sentenceThe tailorrepaints onephoto of meJev is the only part making choices
Where the time goesThe ear · speech to wordsinstant, in the browserThe brain · Jev decides740 ms medianThe tailor · new pictureabout 10 seconds, first timeThe thinking is the fast part
§5One sentence, seven decisions

It asks Jev seven questions at the same time

Seven questions, one callWhat kind of request is this?Which top?Which layer?Which bottoms?Which shoes?Which extras?Which room?One sentence"I'm giving akeynote tomorrow"All seven answers come back together
the real app on my machine · real Jev calls · recorded as it ran

What you're watching: a keynote request fills every slot at once, then "lose the jacket" changes one slot and leaves the other six alone.

I never clicked a single piece of clothing. I only said where I was going.
§6The three kinds of answer

A pick, a grade, or a yes or no

Everything Jev can sayChoicepick one from a listScoregrade it on a scaleYes or nowith how likely yes isThe mirror uses two of the three
§7 Want this wired into your business?

Get the AI agent system I run, with Jev decisions built in.

The mirror is the fun version. The same decision step sorts an inbox, scores leads and hands work to the right agent. Inside the AI Profit Boardroom you get the whole system it plugs into.

The Agent OS — one place to plug in your Claude, your Hermes and your OpenClaw, and add your own workflows on top
The zip file, the 30-day roadmap and a video walkthrough for setting it up, with daily updates as new versions ship
Coaching calls where we go deep on Jev: which decisions to hand over first, and how to read the confidence numbers
Daily tutorials on using Jev-style decisions to save time, sort leads faster and get more customers
3,900+ business owners, some who had never used AI before they joined, and members already building with Jev
A member map so you can find people near you doing the same thing, and someone is always online
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description · used in 38 countries
§8Every answer comes with a number

It tells you how sure it is

"I'm giving a keynote tomorrow" — what kind of request?whole outfit84%no styling change16%scene only0%specific item0%Real numbers from the run you just watched
Same request — which layer?red stage blazer81%keep current layer18%High number: act. Close call: ask.
§9Knowing when to do nothing

I said ten things that had nothing to do with clothes

the real app on my machine · real Jev calls · recorded as it ran

What you're watching: "did I lock the front door" and "how many members joined this week" — the mirror holds still. Then "swap the shoes for the chelsea boots" — it acts, and changes one thing.

My 40-line test40 / 40right calls: act or hold30real requests · all acted on10not requests · all heldA mirror that changes your outfit when you ask about your keys is broken
§10The mistake that made it ignore me

"Heading to the gym" did nothing

One sentence changed the whole mirrorFIRST WORDINGignored me"They want a completelook chosen for them"I said: heading to the gymJev: no change · 80% sureSECOND WORDINGdressed me"They mention where they'regoing. Dress them for it, evenif they never say dress me"Jev: whole outfit · 90% sureJev reads what you wrote. It can't guess what you meant.
Wrong: “I'd need to be technical to build something like this.”
Right: The part that made the mirror smart or stupid was one plain English sentence. The skill is saying clearly what you mean.
Jev can't guess what you meant. It reads what you wrote.
§11It grades its own outfit

A bad grade triggers a second try

the real app on my machine · real Jev calls · recorded as it ran

What you're watching: a podcast request. The first outfit scored 1.9 out of 3, so the mirror re-picked the bottoms and the new look scored 2.7. It says so out loud.

Jev grading Jevpodcast · first try (test run)1.99 / 3podcast · after the fix2.79 / 3wedding · first try1.72 / 3wedding · after the fix2.37 / 3shorts in a boardroom0.02 / 33 of 30 requests fixed themselves before I saw them
“My first pick only scored 1.9, so I swapped the bottoms.”
§12The honest speed numbers

The decision is instant. The picture is the slow part.

the real app on my machine · real Jev calls · recorded as it ran

What you're watching: a brand-new request the mirror has never drawn. Jev decides in about a second, then the scan line runs while the new picture is made — 9.8 seconds here.

Two speedsFIRST TIME~10 secJev decides in under a secondThe image model repaintsthe photo: about 10 secondsCosts about 4 centsEVERY TIME AFTERinstantJev decides in under a secondThe look is already savedThe swap is a crossfadeCosts nothing extraSame pose every time, so the swap feels live
§13What it cost

334 decisions for under half a cent

The whole 40-line test334real Jev decisions in my test$0.0044total Jev cost for all of them$0.04per new picture · one timeJev charges for what you send in. What comes back is not charged.
Wrong: “Running AI agents all day burns through tokens and costs a fortune.”
Right: 334 decisions cost $0.0044. The decision layer is the cheapest part of the whole system.
THINKING IT? "Doesn't the Agent OS need a huge token budget to run?"

No. It runs everyday work on free local models, free APIs slot in for more, and for the heavy work it drives the tools you already pay for — your Claude subscription already includes the Claude CLI.

Inside the Boardroom there are full token-saving tutorials too.

§14You stay in charge

Click any piece. Save a look. Compare.

the real app on my machine · real Jev calls · recorded as it ran

What you're watching: saving a look, clicking the leather jacket and the rooftop by hand, re-wearing a saved look, then holding the corner thumbnail to see the original photo.

§15Same machine, serious jobs

What this pattern already does in my business

Real runs from this week, same model, same key60 emails sorted2.5 seconds40 leads scored1.5 seconds1,542 link decisions5.8 secondsAll 6 email missessat below 0.60 sure8 of 8 bad leadsflagged257 pages of my siteabout one centA situation, a question, allowed answers, and a number
A situation, a question, a list of allowed answers, and a number.
§16Is it too early?

Jev is five days old

Five days from release to a working buildSep 15Jev isreleasedSep 20this mirroris workingWhat carries overclear answers,clear thresholdsThe model will change. The skill stays.
Wrong: “I'll wait until all this settles down before I learn it.”
Right: The model will change. Writing clear allowed answers and choosing the number you act on carries over to whatever comes next.
Don't take my word for it

Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries.

Read the 158-page wins doc →
§17The honest limits

What this can't do yet

Six honest limitsIt's a painted photonot a real fittingSure is not rightit can be sure and wrongTies can flip28% vs 28%New looks take ~10 secthen they're savedVoice needs Chrometyping works anywhereJev is in betaexpect changesWorth knowing before you build on it
§18 Your move

Build your own decision layer, with help.

A mirror that picks my shirt is a small thing. The same step, pointed at your inbox, your leads and your website, gives you hours back every week.

The Agent OS zip — plug in Claude, Hermes and OpenClaw, then add your own workflows like this one
A 30-day roadmap + video walkthrough, with daily updates as new versions ship
Four coaching calls a week — bring your own Jev setup and ask questions live
Daily tutorials on decision steps that sort leads, route customers and save you time
3,900+ business owners and a member map to find builders near you
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Link in the comments and description · used in 38 countries · new tools added the week they ship