Together with Prof. Bertsimas and Prof. Margonis, we developed a novel matching methodology to improve survival models in biomedicine. I am currently working  on extending a similar approach to other domains where the identification of tail subgroups is crucial, such as high-risk insurance enrollees. Here is the link to our manuscript. 

“We explore whether survival model performance in underrepresented high- and lowrisk subgroups—regions of the prognostic spectrum where clinical decisions are most

consequential—can be improved through targeted restructuring of the training dataset.

Rather than modifying model architecture, we propose a novel risk-stratified sampling

framework that addresses imbalances in prognostic subgroup density to support more

reliable learning across the full range of risk.

Our findings suggest that survival model performance in observational oncology

cohorts can be meaningfully improved through targeted rebalancing of the training data

across prognostic risk strata."

In collaboration with cardiologists at Hartford Healthcare, we propose a multimodal machine learning framework combining 12-lead ECG time-series features with electronic health record (EHR) data to classify left ventricular ejection fraction (LVEF)  into 4 clinically relevant categories: normal (>50%), mildly reduced (40-50%), moderately reduced (30-40%), and severely reduced (<30%). 

The LVEF serves as a critical indicator of cardiac function, traditionally assessed through echocardiography—a resource-intensive and time-consuming procedure. Our findings suggest that ECG-derived features may offer a viable alternative for a granular multi-class LVEF classification, potentially reducing the need for echocardiography.  

Currently in the piloting phase, we hope that a real-time prototype of our model deployed within HHC's EPIC database system will provide valuable insights into practical implementation challenges and clinical utility in routine practice. 

This project was awarded the bronze medal at the Yale Health AI Championship in May 2025.

With the guidance of orthopedic surgeons at Hartford HealthCare, the largest hospital network in Connecticut, we are developing machine learning models trained on patient data of different modalities to improve quality of care for hip fragility fracture patients. In particular, we leverage tabular data (laboratory, demographics...), CT scans, and clinical notes to predict 3 targets as identified by our clinical collaborators: readmission after surgery, post-operative complications within 3 days of surgery, and time to surgery after admission. 

In my final-year research project in Oxford, I have developed a mixed-integer programming solver in Julia based on Clarabel.jl as the sub-solver. I was supervised by Prof. Paul Goulart and together with one of his Ph.D. students, we published a paper to the IEEE LCSS journal and CDC conference 2023. My own work was submitted in the form of a 50-page Master thesis by the end of the academic year.

I implemented my own branch-and-bound algorithm and integrated it with a method for early termination of interior-solves, which became the focus of the aforementioned publication. I finally tested and evaluated the solver on standard benchmark problems such as mixed-integer power converter applications,  portfolio optimization, and problems with cardinality constraints. 

The code for the Clarabel.jl solver can be found here (also source of the cow picture).

During this 10-week internship, I extended a human-robot dialogue framework to help end-users understand robot behaviour. I designed, developed, and evaluated example scenarios in a ROS Gazebo simulation environment where a robot is assigned with pick-and-place or object discovery tasks. 

Through integrating different frameworks from various developers into my own code, I became familiar with a diversity of  tools common to robot applications, e.g., the planning framework ROSPlan, a database logging system, and a human-robot question-and-answering system.  Towards the end of the internship, I was collaborating with my supervisor Dr. Lars Kunze and his student on a paper which we submitted to the ICRA2023 conference, entitled “Embodied Question Answering over Knowledge Graphs capturing Robot Observations and PDDL Plans” .

Find my code here and here.

As a team of 3 engineering students, we designed the hardware, software as well as our business model for a hospital service robot. Our robot aims to help reduce  medication waste in UK hospitals and increase staff's with-patient time through relieving nurses of redundant background tasks such as medicine delivery. 

The engineering design (on-paper only) includes everything from ROS control and actuation, to navigation and a fleet management system. We then created a full business plan with our value proposition, market research, marketing and sales strategy, and financial forecasts for the robot as a product and service. 

This project was captured in a combined 90-page report which won us the 3YP Engineers in Business Competition Prize 2022. As one of the 10 finalist teams in the nationwide EIBF Champion of Champions Competition 2022, we received specialist coaching at the Royal Academy of Engineering and pitched our robot to 4 judges in the entrepreneurship industry. This truly rewarding experience demonstrated to me the importance of establishing a sound value proposition before diving straight into an engineering solution.