Custom AI models, trained on your data
Off-the-shelf networks were trained on other people's samples. When they miss your membranes, your filaments or your particles, we fine-tune an existing model or train a new one on your data, and hand you the model, the scripts to run it, and the documentation to keep using it without us.
Work is done by Ian Cooney, Ph.D. (Stanford; first-author Science and Nature Communications), not routed to anonymous contractors.
Not sure whether you need fine-tuning or a new model? Describe the task and we'll tell you which it is, free, before quoting.
- Custom AI model, trained on your data$20,000
One price for the outcome: a model that works on your data, whether that means adapting a published network or designing one from scratch. Delivered trained, validated, wired into your pipeline and documented, and you own it outright, with no licence and no per-use fees.
Fixed price agreed before any work starts. If the job turns out to be smaller than quoted, the quote goes down.
Founding-client rate: the first three engagements are $10,000, half of list, in exchange for a named case study once you are happy with the result.
What we build and train
Segmentation models
Membranes, organelles, filaments and other cellular features in tomograms, for samples where the general-purpose tools miss your feature or hallucinate ones that are not there. Trained to your annotations, validated against tomograms the model never saw.
Particle picking and classification
Pickers trained on your particle in your ice, for cases where template matching and the stock deep-learning pickers drown you in false positives or miss rare views. Includes classifiers that sort picks before they hit your averaging pipeline.
Denoising and contrast restoration
Models tuned to your imaging conditions rather than a published benchmark, so downstream segmentation and picking work on the restored volumes instead of fighting them.
Fine-tuning versus training from scratch
When a published model nearly fits, we adapt it to your data; when nothing fits, we design and train from scratch. You are paying for the outcome either way, so which method your task needs is our problem, and free scoping tells you honestly what we would do.
Annotation strategy
Most projects need far less labelled data than groups expect. We design the minimal annotation that gets a usable model, tell you exactly what to label and how, and can do part of the labelling ourselves.
Deployment on your hardware
The model is delivered running on your workstation or cluster and wired into your existing processing workflow, not left as a repository you have to figure out. Your students can run it on the next dataset without us.
Why this is worth paying for
Hand-segmenting a tomogram takes days, and a dataset takes months of a student's time that produces exactly one paper's worth of annotations. A trained model does the same dataset in hours, does the next dataset too, and applies the same criteria to every volume, which is what makes the quantitative claims in your paper defensible to a reviewer.
The alternative most groups try first is months of a postdoc fighting a published tool that was never trained on anything like their sample. That time costs more than this service does, and it usually ends here anyway.
Describe the task
What you are trying to detect or clean up, roughly how much data you have, and what you have already tried. You get a fixed price back.
Questions
What do we actually receive?
The trained model, the scripts to run it on new data, documentation of how it was trained and how to retrain it, and an example run over your own dataset. You own all of it outright: no licence, no per-use fees, and nothing that stops you publishing with it or retraining it as your data evolves.
How much annotated data do we need?
Usually less than you expect, and fine-tuning an existing model needs far less than training from scratch. Send a description of the task and what data you have; part of scoping is telling you exactly how much annotation is needed and who does it, before any money changes hands.
What does it cost?
Custom AI model, trained on your data, $20,000. Fixed price agreed before any work starts. If the job turns out to be smaller than quoted, the quote goes down. Founding-client rate: the first three engagements are $10,000, half of list, in exchange for a named case study once you are happy with the result. Describe the task and the data and you will get a number, not an hourly estimate that grows.
Does our data have to leave our machines?
No. Training can run on your own workstation, cluster or cloud account through the same temporary remote access we use for software installation, so unpublished data never leaves your control. Where you prefer we run it on our side, data is deleted after delivery.
A published tool already does this. Why pay for a custom one?
If the published tool works on your data, use it, and if scoping shows that it would, we will say so instead of quoting. The custom work exists for the common case where the published model was trained on samples unlike yours and misses your features. Fine-tuning it on a small set of your annotations is usually what fixes that.
Can it plug into our existing processing pipeline?
Yes. Delivery includes integration with the workflow you already run, whether that is RELION, cryoSPARC, IMOD-based tomography processing or your own scripts, so the model becomes a step in your pipeline rather than a separate thing someone has to remember to run.