Point your camera at it. Know what it is.

Most people wait months on a rash, guess at the cause, and buy three products that were never going to work. A dermatology appointment is weeks out. A photo takes a second — and the right active ingredient is usually the whole treatment.

Six conditions, three specialist models each, one photo, under a second. In private beta while we make it hold up on ordinary phone photos — nothing is stored either way.

Private beta

Get early access to the scan

The scan is open to testers rather than the public while we retrain. Our models are accurate on the images they were trained on, but a measurable share of that comes from how those photos were taken rather than from the skin itself — so we are not willing to hand you a confident answer from a phone photo yet. Leave your email and we will open it up to you as soon as it holds up on real photographs.

What beta testers get

The full scan: a ranked differential across six conditions, with every specialist score shown.
Guided capture that tells you when the patch is in focus and filling the frame.
A direct line to us on anything it gets wrong — that is the point of the beta.

How it works

Reusable pipelines, not one fixed model

Diagnostic AI is trained on the diseases we already know about. When something new spreads, that is exactly the wrong shape: the whole model has to be retrained before it can see the thing everyone is looking for. Dermalyze breaks the problem into small, specialised models that snap together — so a new condition is an addition to the system rather than a rebuild of it.

Specialists, not one big model

One multiclass model reads features off the photo and shortlists what it could be. Each shortlisted condition then gets its own three small specialists — one reads texture, one reads elevation, one reads colour. Six conditions today, so eighteen specialists sitting behind that first model.

Roughly 800 ms per image

The whole ensemble is light enough to answer while you are still holding the phone up, and cheap enough to run on every photo rather than a triaged few.

Built for what shows up next

Adding a condition means training three small specialists and updating the multiclass model that shortlists — not retraining a system end to end. Whole-image models have to be rebuilt when an outbreak arrives; ours gets extended.

How the scan decides. The multiclass model shortlists the three to five likeliest conditions. Each one's own panel of specialists then scores the same photo, and the candidate its panel backs most strongly is the answer. Every number is shown under “How the scan reached this” — we would rather you could argue with the result than trust it blindly.

What we do

Four customers, one ensemble

The same ensemble serves four very different customers. The models are the product — the page you are on is the smallest version of it.

The consumer scan

The widget on this page. A photo in, a likely condition out, and the active ingredients that actually treat it — ranked on medical fit for your result.

Clinic decision support

Shadow mode first: the model reads alongside the clinician and is scored against them, with nothing riding on it. We will not sell clinical software before it is cleared.

Pharma and CRO image scoring

Consistent PASI, EASI and IGA reads across sites and timepoints, so a trial's severity scores do not drift with whoever happened to be scoring that week.

Research datasets

De-identified, IRB-governed, and deliberately diverse across skin tones — the gap that keeps dermatology models failing on the patients they were never shown.

Where we are

Honest about the stage we are at

Now
Research and benchmarking
Under IRB, measured against DDI, SCIN and PASSION rather than our own held-out split.
Next
Shadow-mode clinical tryout
Reading alongside clinicians to build real-world evidence. No device claim, no FDA application yet.
After
Trials, then clinics
Image scoring for pharma and CROs, and clinic subscriptions only once we are authorised to sell them.

The team

Who is building this

A small team out of machine learning research, public health, clinical medicine and quantitative modelling:

Neelesh
- Junior in Computer Science at UT Austin, graduating May 2028; Eagle Scout
- Distinguished Winner, Apple Swift Student Challenge 2025, for a SwiftUI game about real-world volunteering
- Started Dermalyze in 2024 as a fine-tuned ResNet at 80% on four conditions; rebuilt it into the multiclass-plus-specialists stack it runs on now
- Lead Full Stack Developer at AiME Technologies since 2024 — the product plus the Proxmox and Cloudflare infrastructure behind it, at 95% uptime
- Telemetry engineer on Longhorn Racing Electric: Kafka and Protobuf ingest, IO overhead down 20%
Jai
- Paid researcher at The Scripps Research Institute, wet-lab and computational, on protein–ligand interactions and dynamics
- Collaborator at the Harvard T.H. Chan School of Public Health on diabetes comprehension and prevention
- Built predictive models for UC San Diego on projects from hundreds of thousands to millions of dollars
Adway
- Undergraduate researcher at UT Southwestern Medical Center
- Built a deep-learning tool (PySide6, OpenCV, YOLO) for patient neutrophil migration — 700% faster analysis, and $3,000 secured to lead the project
- Built MetaboAtlas, comparing drug and endogenous metabolite binding across 1.1 million cells; co-authored manuscript pending
- Certified EMT with 200 hours of clinical experience

Funded, and running on our own metal

We were funded for roughly $5,000 of hardware to start: an RTX 3090, an RTX 3080 and a 72-core, 128 GB Dell PowerEdge server. Every model is trained and served on machines we own, so a full panel costs us electricity rather than cloud credits — which is what lets the scan stay free while it earns its keep.

Get in touch

Clinics running a shadow-mode pilot, sponsors and CROs with images to score, and researchers who need a benchmark set — that is who we want to hear from.

Terms, in plain words

Dermalyze returns a statistical suggestion from an image. It is not a diagnosis, it is not a substitute for a clinician, and it cannot rule anything out — including skin cancer, which it does not screen for. By ticking the box you confirm the photo is yours to submit. Your image is sent to our inference service to produce the result and is not published, sold, or used to train anything. Product suggestions are ranked on medical fit for your result and nothing else; where a link earns us a commission it does not move anything up the list.