The research behind a trustworthy skin AI
Accuracy claims mean nothing without methodology. Here's how we build, validate, and govern the models behind every Tincyglo scan.
97.4%
Average marker accuracy
1.2M+
Labeled training images
6
Fitzpatrick skin-tone categories covered
Quarterly
Model bias re-evaluation cadence
What we study, and how we apply it
Skin AI
Purpose-built models for skin, distinct from general-purpose facial recognition or beauty filters.
Dermatology AI
Trained in collaboration with dermatology-reviewed imagery, not just consumer selfies.
Datasets
Curated across skin tones, ages, and genders, with continuous expansion to close representation gaps.
Accuracy
Every marker is benchmarked independently — we publish per-marker accuracy, not one blended number.
Model Validation
Held-out test sets and dermatologist review gate every model before it reaches production.
Research Process
A structured pipeline from data collection to deployment, documented and repeatable.
Ethics
An internal review process evaluates every model for fairness before it ships.
Responsible AI
Clear disclosure of what the AI can and can't do — including when to see a real dermatologist.
From dataset to deployed model
Every model release follows the same four-stage process before it ever reaches a user's scan.
Data collection
Dermatologist-reviewed and consented imagery, sourced to cover diverse skin tones and ages.
Labeling
Multi-rater annotation with disagreement resolution to reduce single-labeler bias.
Training
Models trained and iterated against held-out validation splits, never the test set.
Validation & review
Bias testing across demographics, then dermatologist review before production release.
What our AI won't do
Tincyglo’s models are built to inform, not diagnose. We’re explicit in-product about that distinction, and every result includes guidance on when a licensed dermatologist should be consulted instead.
We evaluate every model release for performance gaps across skin tone, age, and gender before it ships, and we maintain a standing review process to catch regressions as the model set grows.
Scan images are never used to train shared models without a user’s explicit, revocable opt-in — full detail is available on our Security page.
Interested in a research collaboration?
We partner with academic and clinical researchers on dataset and validation work.