Any dataset you need, we source it. Any model you want, we build it.
Off-the-shelf computer vision isn't enough anymore.
It clears the easy cases. What it leaves behind is the part worth solving, and that needs a model trained on your data, not hand-tuned rules and manual checks.
Free models are built for the average case. Yours has to be built for your work.
Services
Take any one of these on its own, or hand us the whole chain.
Name the data and we'll get it. If it already exists, we license it. If it doesn't, we collect it to your spec, field capture, commissioned shoots, synthetic generation, or a targeted top-up of the exact edge cases your model keeps failing on. If it can be captured or licensed, we can get it.
License a dataset that already exists instead of building one from zero, usually the cheapest way to get moving. Splits are made, class distribution is documented, and the labeling guide ships with it, so you know what each class was supposed to mean.
Your images and video, labeled to a written guide we agree on before anyone draws a box. Ambiguous frames go to a review queue instead of a guess, and the answer gets written into the guide so it never comes up twice.
We train, tune and benchmark on the data we built, then hand over the weights, the export you asked for, and an evaluation report your own team can read without a translation layer.
Any data. Any model. One vendor.
Labels that look fine can still train a bad model. We check for that before you pay to have the whole set labeled.
How a project runs
Nobody should commit a hundred thousand images to a spec they haven't seen applied. So we label a small batch first, for free, and you decide from there.
You send sample data and describe the classes. We write the labeling guide, label 100 images against it, and send them back for you to tear apart.
Annotators work to the approved guide. Anything it doesn't cover goes to a queue and gets a written answer, never a judgement call made alone.
Every batch gets a second pass. We track disagreement rate per annotator and per class, so quality shows up as a number you can argue with.
Dataset in your format, train/val/test splits, class distribution report, and the guide it was labeled against. Retraining stays reproducible.
Pricing
Buy the images, buy them labeled, or buy the model trained on them. Each level includes the one below it. Datasets start at 1,000 images.
Level 1
Quoted by how hard the images are to get. Common subjects are quick and cheap. Anything nobody else has is priced on what it takes to find or shoot it. Either way you see the number before anything starts. Minimum order 1,000 images.
Licensed data, not scraped data.
Free datasets are usually research-only. Everything here is cleared for commercial training with the provenance on record.
Level 2
Quoted on the work and the imagery together. Tags are quick, masks are slow, how many objects sit in each frame changes everything, and the images themselves cost what they cost. Common subjects are cheap, rare ones are not. You see the number before anything starts.
Licensed for the model you are building. Train it, retrain it, ship it, keep the files. Using the same set for a different product, or passing it on to anyone else, is quoted separately.
Level 3
Option A · we hand you the model
What you get is the model. Finding the data and labeling it are part of the price, but the dataset behind the model stays with us. Each price covers a starter dataset sized for the task, on common imagery. A larger set, or a subject nobody has captured, is priced on top. If you want the dataset too, see Who keeps the data below.
Option B · we set that same model up for you
Same model, delivered as an endpoint instead of a file. We package it as a container that runs inside your own cloud, whether that is AWS, Azure, Google Cloud, Alibaba Cloud or your own servers. Your images never leave it and your team calls it like any other internal service. You keep it running for as long as you need it. Quoted per project.
Beyond both · a deeper version of either one
Production tier
from USD 20,000
Not another kind of model, a deeper version of any of them. A dataset several times larger, more rounds of iteration, and help getting the model running where you actually need it. Scoped and quoted per project.
Questions about the price
These are starting points, not a formula.
What a model costs depends on what you are building, what it has to do once it is running, and how accurate it has to be to be worth running at all. Dataset size, class count, boundary precision and how many rounds of iteration you want all move the number, in either direction.
The trained weights are yours. The dataset behind them is not, by default. If the images were yours to begin with, they stay yours. What the data costs on top depends on what it is. Common imagery is a small addition, rare imagery is not.
Priced differently, and only after a paid feasibility study. Plain accuracy is usually the wrong number anyway: on a line that is 99% good parts, a model that calls everything good scores 99%. Detection is judged on mAP, segmentation on mIoU, keypoints on OKS. Nobody can promise a number on data they have not seen.
All prices in US dollars. Invoices can be issued in another currency on request.
Start a project
Tell us the task, the volume, and the two classes you already know get confused. We'll come back with a quote, or the one question standing in the way of it.
Project brief
Pick whatever applies. We’ll come back with a quote, or the one question standing in the way of it.
A confirmation is on its way. If it is not in your inbox, check your junk folder. We reply within one business day.