Diagnosis App
References from this period: 4 LinkedIn recommendations, full text. They vouch for the role, not for any one claim.
Proof points: 4
| Role | Product | Situation | Task | Action | Result | Year | Skills |
|---|---|---|---|---|---|---|---|
| PM/GM at LEO Innovation Lab | Diagnosis App Prototype smartphone app diagnosing skin conditions with machine learning | A moonshot: could a smartphone camera diagnose skin conditions? A 12-person team of data scientists, ML and computer vision engineers, wired for logic and certainty, in the deep uncertainty of product discovery. His first deep experience leading an ML team. | Lead the product direction, prioritisation and strategy, and get a working prototype that could be evaluated. | Worked with dermatologists to redesign closed-loop diagnosis and the ML training platform. Bootstrapped the ML engine with clinical decision trees. Designed the pipeline for data collection and analysis. Models were evaluated against a set of pre-tested images the dermatologists had all voted on. Not a medical device by design: disclaimers made clear it was not a diagnosis. Evaluated go-to-market and identified seven paths to market. | Over 80% accuracy across three skin conditions. Doubled the labelling output of the in-house image labelling platform. | Nov 2018 - Jan 2020 | AI, machine learning, computer vision, healthcare, discovery, prototyping, growth, launch, team leadership |
| PM/GM at LEO Innovation Lab | BetterVisitBot a pre-briefing chatbot for doctor consultations | Patients sometimes get intimidated by the doctor, and don't get the most out of their consultation, or forget what was said. | Prototype it with GPs. | Hid a dermatologist in an adjacent room to simulate a consultation chatbot. | Right idea, too early. Now it isn't. | 2016 to 2020 | AI, machine learning, computer vision, healthcare, patients, discovery, prototyping, growth, launch, team leadership |
| PM/GM at LEO Innovation Lab | Diagnosis App data pipeline Image labelling platform and data pipeline for the diagnosis app | Labelling by dermatologists was critical for model performance, and a significant bottleneck. | Get more labelled data through the system without lowering its quality. | Interviewed every dermatologist on the labelling work to understand how they actually worked, then redesigned the in-house image labelling platform around that. Designed the pipeline for data collection and analysis to feed it. | Doubled dermatologist labelling throughput. Less of their time on labelling training data, more labelled data for the models. | Nov 2018 - Jan 2020 | AI, machine learning, computer vision, healthcare, discovery, prototyping, growth, launch, team leadership |
| PM/GM at LEO Innovation Lab | Diagnosis App paths to market Go-to-market evaluation for a moonshot ML prototype in dermatology | A functioning prototype, but heavy regulatory barriers and no clear path to market. | Work out how, or whether, a moonshot prototype could reach a market. | Evaluated go-to-market systematically and mapped the options with the team. | Identified seven paths to market, with the trade-offs of each, for the lab to decide on. | 2019 | AI, machine learning, computer vision, regulated, healthcare, discovery, prototyping, growth, launch, team leadership |
Tell me about the diagnosis app
source: Tom's own account
PM/GM at LEO Innovation Lab (Nov 2018 - Jan 2020), leading a 12-person team building a prototype diagnosis app using machine learning. This was my first deep experience with ML teams.
Key achievements:
- Led machine learning and computer vision specialists
- Achieved >80% accuracy across 3 skin conditions, doubled labelling throughput
- Evaluated go-to-market for a moonshot ML prototype; identified seven paths to market
Direct answers
- Led a team of AI engineers all wired for logic and certainty through the deep uncertainty of product discovery, to rethink how a smartphone camera could diagnose skin conditions.
- Worked with derms to redesign closed-loop diagnosis and ML training platform. Bootstrapped ML engine with clinical decision trees. Safe to launch with a path to full automation. Doubled dermatologist throughput.
- Evaluation on the Diagnosis App: a golden image set built from dermatologist consensus; the model was a hybrid of deep learning and the diagnostic manual (clinical decision trees); guardrails were not-a-diagnosis disclaimers; non-SaMD by design, with a path to full automation; >80% accuracy across 3 skin conditions and labelling throughput doubled[LEO.R3].
source: Tom's own account, direct answers
Related
- AI: tell me about your AI experience THEME.AI
- Ask: how technical are you TQ.TECHNICAL
- LEO Innovation Lab: tell me about LEO LEO