Projects



ML for Home Health Monitoring

I built a radar-based remote patient monitoring system for people with neurodegenerative conditions living at home. I started building the core technology during my PhD, then grew and led a team of engineers and scientists to translate and commercialize our work. I led engineering across the full stack — custom hardware, embedded systems, edge ML models, user-facing software — ran trials with hundreds of patients and clinicians, and drove the work from research into commercial deployment across senior care facilities and UK NHS hospital-at-home programs.

Imperial College London • UK Dementia Research Institute • UK National Health Service

Problem

Most people with dementia and movement disorders live at home, often alone. Falls, infections, and other clinical deteriorations are commonplace, and by the time anyone notices, the patient is in the emergency room. There's no visibility into what's happening between sporadic clinic visits. To catch the early warning signs in time, patients and their healthcare providers need continuous health monitoring systems that measure what matters: vital signs, movement and sleep patterns, and behavior. Nothing on the market works for this population. Wearables offer limited data and have poor adherence, and would you want a camera in your bedroom?

Approach

We built the full stack from the ground up. We produced award-winning R&D, executed validation trials with hundreds of patients, and secured multiple commercial deployments. We built custom radar sensors that capture movement, vital signs, sleep, and behavioral patterns with privacy-by-design from day one. I led engineering end to end: hardware design, embedded ML models trained on continuous physiological data to detect early signatures of deterioration including falls and infections, IoT infrastructure for multi-sensor home deployments, and software that surfaces actionable insights to care teams early enough to intervene. Operating in high-stakes healthcare meant every data point had to be clinically reliable, every device we manufactured went through extensive QA, and our users — patients, caregivers, clinicians — were in the loop at every stage informing design decisions. Most solutions we saw tried to retrofit generic consumer tech to this population. We built for it.

Outputs

Designed and manufactured fully compliant home-ready remote patient monitoring system, published award-winning research, secured patents, and won commercial deployments in senior care facilities and the UK NHS's flagship hospital-at-home program for patients with dementia and movement disorders.

Audio courtesy of American Cleft Palate-Craniofacial Association

ML for Cleft Speech

I led a multi-center research project across Brazil, Sweden, and the UK to build a language-agnostic ML system that automatically detects, grades, and monitors speech abnormalities caused by cleft lip and palate — replacing subjective clinical assessments with objective, automated measurement.

Great Ormond Street Hospital • University of Edinburgh • Karolinska Institute • University of São Paulo

Problem

Cleft lip and palate affect 1 in 700 births globally, often causing debilitating speech abnormalities despite surgical correction. Assessment depends on subjective listening by specialist speech therapists — who are scarce everywhere and virtually nonexistent in most of the world. Assessments are inconsistent across clinicians, incomparable across languages, and too infrequent to meaningfully track treatment progress. No objective, automated tool existed that worked across languages.

Approach

I built the ML system from scratch — crowdsourced clinical training data from partner institutions, designed a novel synthetic data generation method to augment it, architected and trained the model, and built a mobile application as the deployment medium. A user records themselves speaking and the on-device model outputs objective, quantified measurements, enabling longitudinal tracking of speech outcomes for the first time. Most excitingly, the model generalized across languages. To test and validate the approach against experts, I assembled a consortium of several of the world's leading speech and language therapists, from Great Ormond Street Hospital, University of Edinburgh, Karolinska Institute, and University of São Paulo. Many other engineers and clinicians have continued to contribute to the project which is undergoing trials in Brazil, Sweden, and the UK.

Outputs

Working prototype deployed across hospital systems in three countries. First language-agnostic ML system for cleft speech assessment. This work demonstrated that automated speech analysis could provide consistent, objective measurements across different languages and healthcare systems.