Triage: AI dermatology in a million pockets

Context

Triage Technologies built one of the first consumer AI dermatology products: point your phone at a skin condition, get ranked predictions across 588 skin diseases in 133 categories. In 2017, this meant solving problems the industry now considers standard practice, data flywheels, human-in-the-loop labeling, eval-driven development, years before they had names.

What I built

• Pioneered a 0→1 AI app used by 1M+ consumers via product-led growth, reaching 90%+ top-5 prediction accuracy (from a starting point of 17%).
• Built the largest real-world skin-disease dataset: grew labeled data from ~10k messy images to 500k+ catalogued images with a disease ontology graph and rich metadata; created 200+ offline experiment datasets.
• Invented the labeling flywheel: recruited, vetted, and trained a global network of 50+ dermatologists on a proprietary labeling product powered by a consensus algorithm I built, measurably lifting label quality and model performance, layered with 100k+ crowd-sourced labels and cutting annotation costs 80%.
• Ran eval-driven development before it was standard: designed and ran clinical studies benchmarking the classifier against general physicians and dermatologists (it outperformed average clinician sensitivity and specificity on high-risk lesions), plus eval datasets reused hundreds of times across dozens of algorithms.
• Cut output errors 97% with ML integrity features, content-policy filtering and label-sanity models.
• Launched Triage Clinical, securing pilots with Stanford, UCLA, and Memorial Sloan Kettering; main author of the 2018 FDA pre-submission for AI-based patient decision support, written with Apple's former head of regulatory compliance, successfully arguing consumer AI skin screening from high-risk to low-risk classification.
• Co-authored "Q&A-informed image-based deep learning for skin disease recognition in the real world" (SOCML 2018). The technology was later licensed by MyFiziq (ASX:MYQ) for US$3M.

Skills in play