AI/ML Projects

ClariCare+

AI-Powered Health Understanding and Remote Patient Monitoring

ClariCare+ is an AI-driven healthcare platform designed to bridge gaps in health literacy and remote patient monitoring by transforming complex medical information into accessible, actionable insights for both patients and clinicians. The system emphasizes longitudinal understanding of patient health rather than isolated measurements, with the goal of improving patient comprehension, reducing preventable readmissions, and supporting informed clinical decision-making.

The platform addresses widespread challenges in health literacy by translating clinician-provided medical instructions and terminology into clear, plain-language explanations that are personalized using a patient’s historical health data.
ClariCare+ integrates physiological data from consumer-grade wearable devices, including Apple Watch and Samsung Health, enabling continuous monitoring and trend-based analysis of vital signs over time.
A dual-dashboard architecture separates patient-facing and clinician-facing views, allowing patients to receive simplified summaries while clinicians access baseline-aware interpretations, deviation detection, and contextual alerts.
AI-based summarization and trend interpretation models are used to identify abnormal patterns relative to individual patient baselines, rather than relying on static population-wide thresholds.

A quick demo video of the platform in action is shown below.

Demo Video

Adversarial Transfer Learning Toolkit

Practical Robustness Evaluation for Vision Models

This project presents a lightweight adversarial robustness evaluation toolkit designed for realistic black-box machine learning scenarios. A Streamlit-based interface enables users to upload models and datasets, visualize adversarial examples alongside clean inputs, and export evaluation results for further analysis. The toolkit can be viewed here.

The toolkit emphasizes zero-query, transfer-based adversarial attacks, enabling robustness evaluation when only surrogate models or pretrained architectures are available.
It includes an implementation of the Sine Attack, a frequency-domain adversarial method that parameterizes perturbations using low-dimensional sinusoidal functions, exploiting structured vulnerabilities in convolutional neural networks.
A cross-domain adversarial generation pipeline allows perturbations trained on one dataset or model family to transfer effectively to unseen architectures and domains.
The system reports comprehensive evaluation metrics, including attack success rates, accuracy degradation, perturbation norm statistics, and perceptual similarity measures.

A quick demo video of the toolkit in action is shown below.

Demo Video