Everything below began as somebody's Monday morning: a job done by hand a hundred times a month. You do not need to know which of these you want, or whether any of it applies to you. That is what the first conversation is for.
Start small
You do not have to decide on a project to talk to us. Tell us about the job that eats your week and we will tell you whether it is worth building, or whether there is a simpler fix that is not AI at all. That conversation is free and it does not turn into a pitch.
If it looks worth doing, the next step is still small. We take a real slice of the work, fifty documents, twenty photographs, an hour of audio, and run it on your own material for a fixed fee agreed up front. You see what the output actually looks like, and what it gets wrong, before anyone commits to building the whole thing.
If you can describe it
The menu below is what comes up most, not the edge of what is possible. Nearly all of this work turns on the same thing: once a job can be described clearly, there is almost always a way to build toward it.
So the useful question is not whether a tool for it already exists. It is what you would do with one. Bring us the awkward idea, the one you assumed was too specific to be worth asking about. Those are usually the ones worth building.
Say the problem, find the page
Describe the job in your own words and this finds the matching page on this site. It retrieves, it does not generate: it can only point at pages that exist, so it cannot make things up. First search fetches the matcher, about 28 MB, then it is cached.
What we build
Custom models
Brand models
Generate images that are already on brand, so the review is a yes instead of a list of corrections. The model learns your system, not a generic house style.
Typical jobs
Campaign variations for six markets, without a shoot in each one
A year of social imagery that still looks like the same brand in December
You get
The trained model, as weights or as private access, depending on what the base model's licence allows. We tell you which before you commit, not after. An eval sheet comes with it, showing where the model holds and where it drifts.
To begin
A finished brand system: marks, palette, type, and the written rules. If you do not have one yet, that is the studio half of this business and we can build it first. A model can only learn a standard that already exists.
Task models
Take one job you do constantly and hand it to a model tuned for exactly that. Narrow beats general: a small model doing one thing well will beat a large one guessing.
Typical jobs
Cut-outs and shadows on every new product photo, to the same standard
Grading or colour-matching a shoot to the way the studio already does it
You get
The model, the eval set it was measured against, and the score it got. If it does not beat what you are doing now, we say so and you do not ship it.
To begin
Worked examples of the task, done the way you want it done. A few hundred is the usual shape. Send what you already have and we will tell you whether it is enough before you go collecting more.
Text at volume
Copy at volume
Fill a catalogue nobody has time to write by hand. Every line written against your voice document, so the thousandth description sounds like the first.
Typical jobs
Four thousand SKUs relisted on a new platform, no copywriter booked
A seasonal refresh that still sounds like last season
You get
The copy, in the format your store imports, plus the voice document it was written against so the next batch matches this one.
To begin
Your product data, and a voice document if you have one. If you do not, we write one with you first. It is the difference between copy that sounds like you and copy that sounds like anyone.
Voice and audio
Transcription
Turn a shelf of recordings into text you can search. Speakers kept apart, and the names your trade actually uses spelled the way you spell them.
Typical jobs
A back catalogue of customer calls, made searchable for the first time
Interview footage turned into pull quotes without scrubbing the timeline
You get
Timestamped text, speaker-separated, and a correction pass over the names and trade terms a general model reliably gets wrong.
To begin
The audio, and a list of the proper nouns in it: product names, people, places. Ten minutes spent writing that list is what turns a rough transcript into one you can rely on.
Voice and narration
Narrate a library you could never justify booking a booth for. Re-record a changed line months later and it still matches the rest.
Typical jobs
Forty product videos that each need thirty seconds of voice-over
Training material that gets revised every quarter
You get
The audio, and the script exactly as recorded, so a re-record months later still matches.
To begin
A script. If the voice belongs to a real person, we need their written consent first. We do not clone a voice without it, including yours.
Data and automation
Document pipelines
Stop retyping what arrives as paper. Documents go in, structured data comes out, and you get told when one does not fit rather than finding out later.
Typical jobs
Supplier invoices into the accounting system without a data-entry hire
Spec sheets from thirty manufacturers, all landing in one format
You get
A pipeline that runs on your schedule, its accuracy measured against your own documents, and the cases where it fails written down rather than discovered later.
To begin
Real documents, and please include the messy ones. Tidy samples give a flattering accuracy number that will not survive your actual inbox, and we would rather show you the true one.
Automations
Give the job you do a hundred times a month to something that does not get bored or skip a step on a Friday. You keep the log and the off switch.
Typical jobs
Resizing and exporting one artwork into every placement it has to fill
The weekly report somebody currently rebuilds by hand every Monday
You get
The automation, a log of everything it did, and an obvious way to stop it. Anything that runs unattended needs all three.
To begin
To watch you do the task once, start to finish, and to hear what currently happens when it goes wrong. Half an hour on a call is usually enough to see it.
Search over your own material
Ask a question and get the answer out of your own documents, with the source attached. When the answer is not in there, it says so instead of inventing one.
Typical jobs
Ten years of specs and manuals your team currently asks one person about
Contracts and quotes, searchable by what they say rather than what they are named
You get
Search across your material, answers that cite the document they came from, and a straight 'not in here' when there is no answer rather than an invented one.
To begin
The documents, and the rules for who is allowed to see which. If those rules only live in someone's head, we can help you write them down. Permissions are part of the build, not bolted on after.
Customer-facing tools
Preview tools
Let a customer see the thing before they commit to the order. Fewer changed minds after production, and fewer sample runs to get there.
Typical jobs
A configurator that shows a customer their logo on the actual product
Room or shelf views generated from artwork instead of photographed
You get
A tool that runs in a browser, built on your dielines and your artwork, showing only combinations you can actually make.
To begin
The dielines, the material list, and the rules for what combines with what. Those rules are the hard part, so we usually work them out with you rather than waiting for them to arrive.
Bring the problem rather than the product name. A tedious job, described in one sentence, is enough for us to work with.
Where your data lives
Generation and model training run on hardware we own, in our own building. Client material does not leave our machines to reach a third-party API, and it is never training data for anything but your own model.
We check the licence on every model before its output ships. If a licence will not let you use the result the way you intend to, that rules the model out. We find another one.
And when a finished tool goes live, where it runs is up to you: kept on our hardware, moved to yours, or set up in a private cloud through a GPU hosting supplier. You control it wherever it lives.
Start here
Bring us the bottleneck.
Tell us about the job you do a hundred times a month. The first conversation costs nothing but the email.