I serve as the lead researcher on this project, which focuses on developing a multimodal deep learning framework for estimating beat-to-beat systolic blood pressure changes (ΔSBP) in individuals with spinal cord injury (SCI), a population particularly affected by autonomic dysregulation. The proposed model integrates photoplethysmography (PPG), electrocardiography (ECG), and demographic data within a hybrid CNN–BiLSTM–attention architecture. I designed the full pipeline, from preprocessing and time–frequency signal transformation (via scalograms) to model architecture, training, and evaluation.

To address data scarcity in SCI cohorts, I implemented a transfer learning strategy from the large-scale Aurora BP dataset and fine-tuned the model using leave-one-subject-out validation on the SCI data. The approach demonstrated strong generalization and achieved clinically relevant performance, meeting AAMI standards and achieving BHS Grade B. I also incorporated Grad-CAM to enhance interpretability and ensure physiological relevance of predictions. This work represents the first known application of transfer learning from Aurora BP to SCI-specific ΔSBP monitoring and establishes a foundation for deploying real-time, personalized blood pressure monitoring in wearable systems.

MeReader is an innovative, offline, context-aware AI assistant designed to enhance narrative eBook reading by delivering targeted, spoiler-free support precisely aligned with the reader’s current progress. It uniquely combines a retrieval-augmented generation (RAG) framework with semantic vector search, all executed locally on the device, thereby ensuring strict user privacy without reliance on cloud services. This approach addresses key cognitive challenges in digital reading, such as impaired recall and disrupted narrative continuity, through a minimally intrusive, progress-aware assistance model.

As the project supervisor, I guided the development of these novel technical solutions, including the integration of location-bound context filtering and privacy-preserving on-device inference, leading to a state-of-the-art system that advances both AI research and practical reading technologies.

The main objective of the project was to create a noninvasive, continuous blood pressure monitoring solution using AI-based algorithms for spinal cord-injured patients.

I was the main data science researcher who proposed the explainable, transfer learning-based framework for blood pressure estimation. My work covered various algorithm configurations, experimentations, testing the end-to-end architecture.

Unsupervised clustering of semiconductor wafer map defect patterns was the focus of this project, aimed at detecting subtle, high-dimensional failure signatures for yield optimization. The final output was a modular pipeline combining convolutional autoencoders for feature extraction, PCA for dimensionality reduction, and improved deep embedded clustering (IDEC) with KL divergence to learn confident latent representations. The workflow included SHAP, Grad-CAM, and cluster profile visualizations, along with evaluation metrics like entropy, confidence, and semi-supervised scoring.

I was the lead architect and main researcher—responsible for scoping, architecture, implementation, interpretability, experimentation, and stakeholder communication.

I serve as an advisor to NuralUmi, an innovative project exploring the use of AI agents to streamline academic research and writing. The system leverages cutting-edge LLMs and agentic frameworks to assist researchers in transforming complex topics into well-cited, structured drafts. My role includes guiding technical direction, ensuring academic relevance, and supporting the development of intelligent, responsible tools for the future of scholarly work.

In the rapidly evolving semiconductor industry, optimizing production processes and improving yield are critical challenges. This article introduces a flexible, AI-driven framework designed using Design Science Research (DSR) principles, specifically addressing the clustering of wafer map patterns in MEMS. By prioritizing explainability and adaptability, the framework enhances model transparency and user trust, making it applicable across various AI-driven projects. Evaluated through a wafer map clustering case study, the framework demonstrates its effectiveness in delivering precise, interpretable results, paving the way for more reliable and efficient AI applications.

This project deals with improving second-level calibration in sensor manufacturing. To automize and digitize the existing state-of-the-art solution, we proposed a quasi-parallel fail-safe approach. An end-to-end monitoring solution resulted in a 23.8% improvement in calibration time.

I was the main responsible data science researcher to handle the following tasks: research gap analysis, solution design formulation and verification, the main architect of the monitoring design solution, and experimentations.

This project focused on predictive maintenance for MEMS-based inertial sensors using machine learning. We employ a Design Science Research (DSR) approach to develop a process model that enhances manufacturing efficiency and facilitates knowledge reuse.

I led the development of a predictive maintenance system for MEMS inertial sensors using machine learning, following a Design Science Research (DSR) approach. This involved designing and implementing the system, including algorithm selection and integration into manufacturing processes. DSR principles ensured a systematic approach, improving sensor performance and production efficiency.

The project focused on predicting thermal drift-related failures in MEMS-based gyroscope sensor production. We used various ML-based and statistical approaches for smart feature selection to find the optimal set of features to reduce time and effort. Finally, with the reduced set of features and XGBoost algorithm, abnormal drift values on a single wafer could be predicted with a precision of 85%.

I was the lead data scientist for this project. I worked on almost all aspects: technology research, research gap analysis, solution design and experimentation, paper publication

The project integrates Design Science Research (DSR) and Business Process Management (BPM) methodologies to create a comprehensive framework for engineering digital innovation. This framework aims to enhance the creation, transfer, and generalization of digital innovation ideas while optimizing organizational processes through structured and flexible design strategies.

The projects aimed at predicting the movement of asset prices in the stock market. For this work, Convolutional Neural Networks (CNN) and Vision Transformers (VT) were used on three data types: heat maps, correlation metrics, and line plots.

This was a master's thesis work under my direct supervision. My work covered problem formulation, experimental design and in-depth analysis of the final work.

The project revolves around the solution development for the identification and forecasting of traffic conflicts in connected autonomous vehicles (CAVs) using Vehicle-to-Everything (V2X). With the help of ML models, the final conflict model was designed.

This was a master's thesis work under my direct supervision. My work covered problem formulation, experimental design and in-depth analysis of the final work.

The project dealt with the classification of open-source software using ML solutions. A great effort was made to define what can be categorized as “open source software” with almost more than 200 repos.

This was a master's thesis work under my direct supervision. My work covered problem formulation and experimental design. algorithm suggestions and in-depth analysis of the final work.