Your MedTech business generates valuable data.
We help you turn your proprietary scans, images, videos, exams, and more into foundation models that power new business opportunities.
You already own the hard part.
New model architectures are published every week, and anyone can download them. Nobody can download the years of data from your installed base. That archive is the scarce resource, and it’s already on your servers.
What’s usually missing is the know-how. Hiring the right team is slow and expensive, and people who have trained foundation models rarely join device companies to start from scratch. Working with someone who has done it gets you a model first and a trained team along the way.
| Hire a team first | Working with Pretrain Labs | |
|---|---|---|
| First foundation model | A year or more | 12 weeks after the assessment |
| Commitment | $1M to $2M a year in new hires | One fixed-scope project |
| Your current team | Waits for the new hires | Learns by building with us |
| What you keep | Whatever stays when people leave | Weights, code, evaluation harness and runbook |
Foundation models unlock new business opportunities
A foundation model isn’t the product. It’s what makes the next ten products cheaper to build.
The foundation
Your foundation model
One backbone, trained once on your archive and owned by you.
- Scans
- Images
- Video
- Exams
- Outcomes
Faster AI products
One backbone turns each new detection, measurement or scoring model into a small project instead of a new program.
Clearable software
Task models built on the backbone can go through 510(k), with the validation and credibility documentation FDA expects.
New diagnostics from existing images
Opportunistic screening looks for disease the image was never taken for, using images you already capture.
More value from outcomes data
Pair images with what happened next to predict complications and pick the right patients.
A platform others build on
Research groups building on your model produce publications, clinical evidence and demand for your devices.
A credible AI story
Your board and investors see a model and measured results, not a slide.
How we’d work together
Four steps, each worth doing on its own. Most companies start with the assessment.
1
Data assessment
Start here
A paid, fixed-scope review of what you have. I look at volume and modality mix, the labels and outcomes already linked to your images, data rights and consent, and where the compute would run. You come away knowing what a model on your data could do and what it would take.
You get a written assessment, a ranked list of use cases and a build plan with cost and timeline.
2
Foundation model build
12 weeks
A self-supervised backbone trained on your archive, in your cloud, with weights you own. I benchmark it against your current models and against open weights like DINOv2 and DINOv3, so the gain is measured rather than claimed.
- Weeks 1 to 4Data transfer and curation
- Weeks 5 to 6Evaluation harness and open-weight baselines
- Weeks 7 to 10Pretraining runs and ablations
- Weeks 11 to 12Model selection and handoff
You keep the checkpoints, training code, evaluation harness and a runbook.
3
Handoff and first applications
Your team takes over
I work alongside your engineers on the first task models built on the backbone, set up evaluation and regulatory documentation, and make sure they can retrain and extend it without me.
You get your first task models and a team that knows how to build the next ones.
4
Fractional AI lead
Ongoing
A monthly retainer for ongoing guidance on architecture, data, partnerships and hiring, plus a technical voice in front of your board, investors and FDA when you need one.
Engagements are contracted with Pretrain Labs LLC and run in your environment, on your data. Pretrain Labs is independent of Virgo, and client work never uses Virgo data or IP.
An example engagement
A composite example to show the shape of the work, not a client story.
Say you make a clinical photography system used in dermatology practices and psoriasis trials. You have 350,000 imaging sessions, each with standardized photos of the body from several angles, many followed visit by visit through months of treatment, and a small data science team maintaining a few task-specific models.
Before the build
Assessment
Confirm data rights, including consent for trial images, and pull a representative sample. Inventory the labels you already have, like clinician-scored PASI and body surface area estimates, and pick the first tasks worth benchmarking.
Weeks 1 to 12
Build
Pretrain on single photos and on full multi-angle sessions. Benchmark plaque segmentation and severity scoring against your current model and open weights, scored on your dermatologists’ gold-standard labels and checked across skin tones.
After week 12
Applications
Fine-tune the first task models with your team, starting with automated PASI scoring for trial endpoints, predicting treatment response from the first few visits, and flagging psoriatic arthritis risk from nail changes in the same photos.
Track record
- EndoDINO. A foundation model for GI endoscopy, pretrained on 10 million frames curated from 130,037 videos, with vision transformers from 86 million to 1 billion parameters. It set state of the art on landmark classification, polyp segmentation and Mayo scoring. Read the paperMeta’s write-up
- A four million video archive. At Virgo I built the capture hardware, cloud and EHR integration behind the largest known GI endoscopy video dataset, eight years in the making.
- Regulatory. Led an FDA ISTAND qualification effort for an AI-derived clinical trial endpoint, and wrote model credibility documentation to FDA’s framework.
- Patents. First-named inventor on two US patents on medical video, US 10,342,410 and US 10,636,518.
- Devices first. I started in medical devices and robotic surgery, and I’ve been building computer vision models since 2015.
Anyone can download an architecture. Nobody can download your archive.
About
I’m Matt Schwartz. I'm a 2x startup founder in the medical AI space (Virgo and Zaniah Health). I first fell in love with machine learning back in 2015. Ever since, I've been pushing the frontier of applying self-supervised learning to medical imaging applications. In medicine, the scarce input is well-curated clinical data.
Pretrain Labs is how I help device and imaging companies put the data they already own to work. I take on a few engagements at a time and work hands-on, in your codebase and your cloud.
- Based in
- Carlsbad, California
- Company
- Founder and CEO, Virgo
- Paper
- EndoDINO, arXiv 2501.05488
- Press
- Meta AI on building EndoDINO with DINOv2