I am currently the lead data scientist at Twin Health, a health-tech startup taking a data & ML-driven approach to treating chronic health conditions. During my time here, I have architected all of our data science infrastructure, implemented the core ML-driven elements of our product offering, built scalable data pipelines, and developed novel applications of ML. Three features involving blood glucose prediction stand out as noteworthy examples. Given our focus on Type II diabetes, understanding the impact of nutrition on BG is of obvious importance. Understanding how a meal will affect a certain patient's BG, and specifically being able to predict that impact in advance, makes our meal planning feature especially powerful. To deliver that predictive capability, I implemented, trained, and deployed a state-of-the-art, patent-pending model for accurately predicting BG response to meal consumption.

We also need to be able to estimate BG over longer horizons. Typically, this kind of monitoring is accomplished via the dreaded finger prick or a continuous glucose monitor. As an alternative to these traditional approaches, I created a model to predict individual's daily average BG levels over long-term horizons of up to 120 days, at a similarly accurate state-of-the art level. Deployed as a microservice, this model drives our VirtualCGM feature.

Finally, while ultra-short and moderate-long term predictions deliver significant value to our patients, these models require a not insignificant amount of effort on the patient's part, especially in terms of meal logging. To alleviate this friction and to enable extremely long-term predictions of our patients' health, I created a model that uses passively collected data from a smart watch to predict a patient's wellbeing in terms of both metabolic (BG) and cardiovascular (blood pressure) health at exceptionally high levels of accuracy. This model powers our Long-Term Monitoring suite of features and is patent pending.