Showroom
What the lab has built
Bring an idea, build the digital co-worker your life needs. Each experiment here is led by a person, and AI agents do the digital work.
About $3,200,000 of work as an AI prices it, across 60 days worked. How both numbers are made.
The lab's own work
-
big-sister
Watch a lab where agents do most of the work, live. In the week to 10 Oct 2026, they made at least 91% of about 1,100 code changes.
$15,000 to have built
See the project
-
Creative studio
Every logo, icon, portrait and pattern made in the lab sits in one place. That is the look of everything you see here.
$15,000 to have built
See the project
-
Ecosystem Radar
Who else is building for people who work with AI agents, and where are the gaps? Checked again every Monday.
$6,500 to have built
See the project
-
Work/Life HQ
The private morning screen for a company run with digital co-workers. What needs your decision sits at the top, and quiet work stays out of the way.
$220,000 to have built
See the project
-
Lab operations
Everything behind the scenes: joining, seats, the chat. Built so that welcoming the next member takes nobody a day.
$1,200,000 to have built
See the project
-
Olaf//2
The academy's first student is its founder's own digital co-worker. You can't build someone else's until you've built your own, failed, and made it better.
$120,000 to have built
See the project
-
Olaf//Desk
Paper in, one checklist out: an assistant for post, bills and deadlines that never pays or sends anything without a yes.
$300,000 to have built
See the project
-
Oovika
Describe an image, a video, or spoken audio on your iPhone. You see the price before anything is sent. Free on the App Store.
$100,000 to have built
See the project
-
OpenSEO
The lab's own SEO desk: one private install, shared by every seat, for the lab's public sites.
$8,200 to have built
See the project
-
Property VR
Building plans plus public Swiss geodata become a 3D model you can walk through in the browser.
Not valued yet
See the project
-
Revenue lab
A place like this does not pay for itself yet. Phase 2 tests the lab's live channels and five digital products against its monthly running cost.
$38,000 to have built
See the project
-
Wingman
Talk with an AI through your glasses. It sees what you see, answers in your ear, and hands bigger jobs to your own agents.
$130,000 to have built
See the project
-
Tiny
Talk to Tiny at home or in the office. Ask a question aloud, hear an answer, or pass a task to the lab’s agents.
$100,000 to have built
See the project
-
tiny-claw
The lab's own agent. Hand it a job and the answer comes back after it does the reading and searching.
$23,000 to have built
See the project
-
tone-deaf
Hand it lyrics and a score, and it sings them: an AI voice on the melody, a synthesized band underneath.
$25,000 to have built
See the project
-
Babel Rebel
Someone speaks; your iPhone says it out loud in your language. In private testing, not open to the public yet.
$88,000 to have built
See the project
-
Worker pods
A private, always-on machine for every lab seat, where the member's digital co-worker runs with the lab's tools and asks for nothing.
$23,000 to have built
See the project
-
This website
This website is written by hand, with no tracking, so every claim on it can be checked by reading it.
$24,000 to have built
See the project
Members' projects
-
Rose Handmade
A shop front for handmade bracelets: the pieces, how a custom order works, and how to ask for one.
$4,000 to have built
See the project
-
NEO//SCAN
A network scanner with a Matrix look. Double-click it, press SCANNEN, and it shows every device on your network and which of its services are open.
$3,000 to have built
See the project
-
Picked by Rose & Olaf
A shop of things Rose and Olaf would buy themselves, and digital products sold on the site.
$21,000 to have built
See the project
-
SignalForge
A research pipeline that ingests market data, trains models on it and tests whether an idea survives out of sample — built to reach a verdict, including 'no'.
$710,000 to have built
See the project
Build yours here
Tell us your idea. One sentence is plenty.
Where these numbers come from
The amounts are rough AI estimates, in US dollars, of what it would cost to have each project made. The full method below explains their limits.
The full method
Older estimates without a manifest (a record of which files the model read) leave it unknown which files they covered. Lab and RevenueLab were priced again after their old estimates were shown to cover the wrong files. Manifests were recorded for those new estimates; they say nothing about the older ones.
With review-contiguous-v1, source files too large to read at once are cut into consecutive stretches of characters. Every character is kept, but code and the context around it can be split at a cut. Estimates for parts can be off in either direction. These are rough AI estimates of what it would cost to commission the work, not quotes and not money earned.
Days worked counts calendar days with at least one saved change, not how long the project has existed, and not how many hours anyone sat at a desk. Three changes in one afternoon count as one day. The total at the top counts a day on which anything here moved, so a day that four projects worked on is still one day.
Lines written counts the text of the project as it stands today. Images, fonts and machine-generated lock files are left out, because their size says nothing about the work.
Design work is counted in pieces instead. A logo has no lines, so counting them would price a year of drawing at nothing. A design project is measured by the finished pieces it delivers, and its days and start date are read from each piece's own history, including the ones still filed with the project they were drawn for, so a piece is dated from when it was made, not from when it was catalogued. The money question changes with it: the model is asked what a studio would charge to deliver this list of pieces, and it is given the list (formats, sizes, what each one is for), not the artwork itself, because a wordmark has the name written in it.
What it would cost to have built is not our figure. An AI model is given the project's source code, a short description of what it does, and its size in files and lines. The project's notes and lock files are left out, and its name, its lead's name, dates written like 2026-10-15 and words like "AI" and "agents" are replaced. Comments in the code can still mention a date written another way, or how long something took. The model keeps its usual tools, but the question tells it to judge only the material it is given, and never says how quickly the work was made. A model told "this took three days" would price three days of work. It is asked how many hours a competent independent developer or small studio would need to build it and at what hourly rate, and the figure is those hours times that rate. It is asked to be conservative and to estimate low when unsure. It can still read some things from the code itself, for example that one person uses it.
A project too large to read at once is read in parts, normally split along folders, with whole files in one part. The range method above handles oversized individual files. Each part is told that it is one part of a larger project and which one, is priced on its own, and the parts are added up. Its page says so.
The estimates move. Asked the same question twice, the model does not give the same figure. So it is asked three times, and the page shows the middle answer, rounded to two digits, with the lowest and highest next to it. A project is priced again only when its code has changed, and then at most once a week; its page says when it was last priced. Until 15 Oct 2026 these figures came from a different method, which was shown a short description and a list of file names instead of the code; they are not comparable.
What it is worth to own is the same model's answer to a different question, and it is usually small. Most of these have no users and earn nothing, however much they cost to build. We show both because showing only the first would be a sales pitch.
These are estimates, and we would rather they were too low than too high. If you think a number is wrong, we would like to know: olaf@paradaice.info.
Measured 16 Nov 2026.
Work