Professional Projects
This page highlights my professional projects across various fields, including artificial intelligence and computer vision. These projects demonstrate my expertise in developing, and bringing to market, innovative AI-driven solutions, from clinical decision support systems to advanced machine learning models for medical diagnostics and imaging.
Xylem Robotics
AggroAI - AI-based pick and place for agriculture
As a consultant, I developed a fully automated pick-and-place robotic system for packaging applications in the agricultural sector. I managed all aspects of data acquisition, curation, and analytics, as well as the AI methodologies, including model architecture design utilizing foundation models, fine-tuning, and validation. I also oversaw the integration of AI software with industrial robotic arms, ensuring seamless system operation and optimal performance.
Elucid Bioimaging
Lead Machine Learning Engineer
I developed advanced imaging algorithms to enhance data pipelines for predicting fractional flow reserve (FFR) from cardiac imaging. This included creating data curation algorithms and implementing image augmentation methodologies to optimize deep learning models for FFR prediction. Additionally, I designed and managed internship projects focused on the application of large language models (LLM) and latent variable models (LVM), which involved fine-tuning and training CLIP-based models, as well as utilizing stable diffusion for both FFR curve prediction and image generation.
AIOptics
Head of AI
I spearheaded the innovation, development, and deployment of AI products, leading the AI and software efforts for the automatic diagnosis of diabetic retinopathy using a handheld camera. This project successfully delivered solutions to key stakeholders. I built, mentored, and managed a team throughout the design, implementation, and validation of the company’s product-critical software, supporting its advancement beyond the clinical pivotal study, while also overseeing the data analytics pipeline.
I implemented software infrastructures, MLOps practices, and algorithmic designs, defining the company’s R&D strategy and future direction, including analytical models, algorithm validation frameworks, and the data cleaning, labeling, and preprocessing pipeline. Additionally, I developed comprehensive software documentation (SRS, SOP, SDD, SDP) in compliance with FDA regulations.
​
My work also resulted in several patented algorithmic innovations, including retinal image analysis using video, image quality assessment for diagnostic devices, a modular architecture for AI-integrated medical devices, and a hierarchical multi-disease detection system leveraging feature disentanglement and co-occurrence exploitation for retinal analysis. On the technical side, I contributed to deep metric learning models, optical disk detection, image quality assessment, deep learning model interactions, data analysis, and automated labeling.
Myocardiak
Founder and CEO
I founded a digital health startup focused on delivering personalized cardiac diagnostics to predict heart failure following myocardial infarction. The core technology generates computer-aided diagnostics by creating patient-specific digital twins to enhance cardiac healthcare outcomes.
​
In my role, I oversaw business development, scientific research, application development, and IT infrastructure, while managing a dynamic team of five. I spearheaded the development of machine learning algorithms and analytics to classify and predict the likelihood of heart failure or other major adverse cardiac events (MACE) in individuals. Additionally, I built the business infrastructure, fostering relationships with clients and investors, which led to significant business traction.
Oxford University (UK)
Senior Research Fellow
I developed machine learning and deep learning models to classify genetic cardiac subgroups using features derived from shape models.
​
My research explored the correlation between phenotypic and genotypic traits in sarcomere gene-positive and gene-negative groups, as well as hypertensive and non-hypertensive normals, utilizing data from the UK BioBank.
University of Nagoya (Japan)
Research Fellow
During my fellowship at the University of Nagoya, I focused on the development of a clinical decision support system for colorectal cancer, leveraging machine learning as part of a national project by the Japan Agency for Medical Research and Development, which successfully obtained PMDA regulatory approval. The system, called "EndoBrain" is now fully operational in Japan.
Additionally, I supervised several master's and PhD students on research projects, including time series analysis of 3D pathological liver meshes and SLAM-based bronchoscope tracking systems.
Oxford University (UK) - Kellogg College
DPhil (PhD) research
My PhD (DPhil) research surrounded the development of advanced tools and methodologies to generate accurate 3D cardiac meshes from Cardiac MRI (CMR) images, addressing spatial misalignment challenges between slices acquired during different breath holds.
My key contributions include an automated image registration framework to correct translational and rotational misalignments using line intensities and local phase vectors, adaptable to various CMR protocols and directly applicable in clinical settings.
Additionally, I created a surface mesh reconstruction method capable of handling sparse, heterogeneous, and non-parallel contours, enabling the generation of patient-specific surface meshes for biophysical simulations, longitudinal studies, or surgical planning. My work also integrates simultaneous alignment and segmentation corrections to produce spatially consistent slice arrangements, fitted to the image data.
I further developed machine learning algorithms, including U-nets, for automatic segmentation of the left and right ventricles from MRI images, enhancing robustness across various cardiac conditions. This included designing data augmentation algorithms using deformations learned from both pathological and normal cohorts.
Altogether, these innovations form a comprehensive solution for generating accurate 3D cardiac models, supporting precision medicine and clinical applications.