October 6, 2026

---

Can an AI system listen to a patient describe their symptoms and determine how urgently they may need medical attention?

My interest in that question began while volunteering at Grady Hospital in Atlanta. I saw how difficult it could sometimes be for patients to communicate what they were experiencing, particularly when English was not their first language. Conversations could slow down or require repetition, not because anyone stopped trying to understand, but because important details could be difficult to communicate across a language barrier.

That experience first led me to study cultural and linguistic differences in AI mental-health chatbots. Presenting that research at the Grady Research Symposium pushed me to think about what could come next: rather than only measuring where existing AI systems struggled, could I build something specifically around the moment when a patient first describes what is wrong? That question became OneVoice Triage, a voice-first, multilingual emergency assessment system designed to turn natural spoken descriptions into structured clinical predictions.

I wanted OneVoice to make initial triage more accessible and inclusive, helping patients explain their symptoms in their preferred language and receive attention that matches their urgency. For doctors and nurses, it could provide a rapid, data-informed assessment of urgency and possible diagnoses, helping them recognize urgent cases sooner and prioritize patients more consistently.

A patient begins by describing symptoms aloud rather than filling out a conventional questionnaire. OneVoice processes the speech, identifies the language, and converts the patient’s description into a standardized representation for analysis. The system then produces two main outputs: an urgency classification (LOW, MEDIUM, HIGH, or CRITICAL) and a ranked list of possible diagnoses.

I designed and developed OneVoice’s processing pipeline using Python, Java, and Kotlin. I prepared and processed the medical data, fine-tuned the Bio_ClinicalBERT-based classifier using DDXPlus and MIMIC-IV-ED, and programmed the Android app’s audio input, language selection, symptom analysis, and results display. I connected speech recognition, language identification, multilingual translation, and medical classification into a working system, then debugged and tested how information moved through each stage. A substantial part of my work involved balancing model performance, response time, and memory use so the complete pipeline could run directly on a phone.

I built OneVoice to support spoken symptom descriptions in nearly 100 languages, with accessibility central to its design. Once installed and set up, it works without an internet connection. Speech recognition, translation, and medical classification run directly on the device, and the patient’s audio stays on the phone. This makes offline operation particularly relevant to rural communities, clinics with limited resources, and disaster areas with unreliable internet or no connectivity.

Testing examined both predictive performance and response time. The system achieved a macro-F1 of 0.74 for urgency classification, with per-class accuracy ranging from 70% to 92%, and the correct urgency level appeared among the top three predictions in 89% of evaluated cases. End-to-end processing latency was approximately 2.2 to 2.9 seconds in testing.

Working on OneVoice also changed how I thought about medical AI. Building a system is only part of the challenge; it also has to be tested against the different ways real people communicate and against the populations it is intended to serve. A model that performs well overall can still fail in important ways when language, context, or patient populations change. To test OneVoice beyond the data, I asked emergency physicians at Grady Hospital and Emory to review the app, and their feedback shaped how I think about where it could fit into real triage workflows.

The contribution of the project is broader than simply predicting a diagnosis. OneVoice investigates whether speech recognition, multilingual language processing, and clinical machine learning can be integrated into one lightweight triage pipeline that assesses both what might be wrong and how urgently the patient may need care.

Ultimately, the project grew from something I observed at Grady into a question I could actually investigate: Can technology preserve more of what a patient is trying to communicate at the very beginning of care? OneVoice is one attempt to explore that question. It is intended as an emergency-assessment research system rather than a replacement for professional diagnosis or clinical judgment.

Click here to download the Logbook

Click here for GitHub

Click here to download the paper

Thanks for reading my blog !

September 30, 2026

About three weeks ago, I started developing MedLens, a Chrome extension designed to make screening research on PubMed faster and more organized. What began as a small project has now become a fully published Chrome extension that anyone can install directly from the Chrome Web Store.

When researchers conduct literature reviews, they may need to work through dozens or even hundreds of PubMed results. Constantly opening articles, returning to search results, and keeping track of which studies should be included or excluded can quickly become inefficient. MedLens was built to simplify that process.

The extension displays article abstracts alongside PubMed search results, allowing users to review studies without repeatedly opening new pages. Researchers can record Include, Maybe, or Exclude decisions, organize exclusion reasons, maintain a screening log, and export their screening records to Excel.

One of my goals was to keep the tool focused on assisting the research workflow rather than replacing researcher judgment. MedLens can provide rule- or keyword-based screening suggestions, but the researcher remains responsible for the final screening decision.

After several weeks of development and testing, MedLens is now available to download from the Chrome Web Store.

For me, this project has been an opportunity to turn a problem I encountered in research into a tool that other researchers can actually use.

Download MedLens on the Chrome Web Store (available as of Sep 30)

July 22, 2026

What if lifesaving medication were available at the press of a button, exactly when someone needed it most? NarConnect is a community-focused public health initiative developed during Yale Young Global Scholars (YYGS) to address opioid overdoses in New Haven, Connecticut. The project proposes installing free, publicly accessible naloxone dispensing stations, called NarcBoxes, alongside educational workshops designed to teach overdose recognition and emergency response.

As Chief of Staff of our seven-member team, I helped develop a comprehensive implementation strategy, including community outreach, strategic placement of NarcBoxes, budgeting, and long-term sustainability planning. Our proposal outlines an initial deployment of 10 stations containing 1,000 naloxone doses, supported by 20 educational workshops. Through NarConnect, we explored how public health, community engagement, and accessible technology can work together to empower bystanders and potentially save lives.NarConnect: Making Lifesaving Naloxone More Accessible

What if lifesaving medication were available at the press of a button, exactly when someone needed it most? NarConnect is a community-focused public health initiative developed during Yale Young Global Scholars (YYGS) to address opioid overdoses. Originally designed around New Haven, Connecticut, the project proposes installing free, publicly accessible naloxone dispensing stations, called NarcBoxes, alongside educational workshops designed to teach overdose recognition and emergency response.

As Chief of Staff of our seven-member team, I helped develop a comprehensive implementation strategy, including community outreach, strategic placement of NarcBoxes, budgeting, and long-term sustainability planning. Our proposal combines immediate interventions with a broader vision for sustainable overdose prevention, emphasizing not only access to lifesaving medication but also education, community involvement, and reducing the stigma surrounding opioid addiction.

In the short term (Years 1–2), our goal is to transform NarConnect from a YYGS capstone proposal into a community-based public health initiative. While New Haven served as the pilot location for our original implementation plan, I hope to explore opportunities to bring NarConnect to Atlanta, Georgia, where I currently live. Building on our initial proposal to establish 10 NarcBoxes containing 1,000 naloxone doses, I aim to identify high-need locations in Atlanta and explore partnerships with local hospitals, public health organizations, community centers, and harm reduction programs. These stations would provide free, anonymous, around-the-clock access to naloxone, particularly in areas where immediate access to emergency resources may be limited. Alongside these installations, our plan includes organizing 20 educational workshops over 18 months, training approximately 500 community members to recognize opioid overdoses, administer naloxone, and respond effectively in emergencies. Through these efforts, I hope to adapt our original New Haven model to Atlanta’s specific community needs and begin building the partnerships necessary to make the initiative a reality.

In the long term (Years 3–5), our vision is to expand NarConnect into a sustainable and potentially replicable public health initiative. Our original proposal outlines expanding the network to 40 NarcBoxes, making 4,000 cumulative naloxone doses available, and training 1,500 residents. Beyond increasing access to medication, we hope to establish lasting partnerships with healthcare institutions, local governments, and nonprofit organizations to secure diversified funding and support the continued maintenance of dispensing stations. Another central goal is to address the stigma surrounding opioid addiction through public awareness campaigns, educational outreach, and stronger connections to treatment and recovery services. We also envision developing an open-source implementation framework that other communities could adapt, allowing the initiative to expand beyond its original pilot setting. Ultimately, our long-term ambition is to contribute to policy changes that encourage accessible naloxone distribution in public spaces and help communities respond more effectively to opioid-related emergencies.

Through NarConnect, we explored how public health, community engagement, and strategic planning can work together to address a pressing healthcare challenge. Although the project began as a collaborative initiative at YYGS, I hope to carry its mission forward by exploring real-world implementation opportunities in Atlanta and, eventually, beyond. More than simply distributing medication, NarConnect represents a vision of communities where ordinary people have the knowledge, resources, and confidence to intervene when every second matters.

Connect. Educate. Save Lives.

Click here to download

June 30, 2026

---

What happens when an AI model performs well overall but struggles with the cases that matter most? This project was developed during an internship at Stanford University under the mentorship of Dr. Pat Kasinpila, where our team investigated the use of machine learning for coronary artery disease (CAD) classification. We compared different modeling approaches and examined not only how accurately AI could classify disease, but also how class imbalance and dataset composition could affect its reliability.

The project focused on three forms of coronary artery disease: obstructive CAD, nonobstructive CAD, and spontaneous coronary artery dissection (SCAD). A convolutional neural network (CNN) was developed to learn visual patterns associated with these different categories, while logistic regression provided a simpler baseline for comparison.

The CNN achieved a mean validation accuracy of 92.13% without class rebalancing, compared with 84.42% for logistic regression. At first, the CNN appeared to be the stronger approach. However, the dataset was not evenly distributed across the disease categories, raising an important question: was high overall accuracy actually reflecting reliable performance across all three conditions?

To investigate this, the CNN was also evaluated after oversampling to address the imbalance between categories. Accuracy fell to 61.67%. Rather than simply treating this as poorer performance, the result highlighted one of the central challenges of medical machine learning: a strong overall accuracy can conceal weaknesses involving less-represented patient groups or disease categories.

The main contribution of the project was therefore not simply determining which model produced the highest accuracy. It demonstrated how class imbalance, rare conditions, dataset composition, and model choice can substantially change how reliable an AI system appears to be.

Ultimately, the project shows why evaluating medical AI requires looking beyond a single performance number. Understanding where a model struggles and why can be just as important as measuring how often it gets the answer right.

Click here to download

Thanks for reading my blog !

May 11, 2026

Can an ICU prediction model estimate several possible patient outcomes without contradicting itself over time? This project develops Joint-CE (Joint Competing-Event), a machine-learning model that treats major ICU outcomes as competing possibilities—meaning the first event to occur determines the patient’s next clinical state. Using MIMIC-IV and eICU-CRD, Joint-CE predicts both what may happen next and when across multiple future horizons. Unlike separate prediction models, it produces one coherent probability distribution across outcomes and time, preventing logically contradictory risk estimates.

The model focuses on patients who have not yet received major organ support and repeatedly updates predictions during their ICU stay. At each landmark, it estimates the first major event that may occur: invasive ventilation, vasopressor support, renal replacement therapy (RRT), combined ventilation and vasopressor support, death, or live discharge. If none occurs within the prediction window, the patient is considered event-free. Predictions are generated for 24-, 48-, 72-, and 120-hour horizons.

The main contribution of the project is its approach to probability coherence. Conventional methods can train a separate model for each outcome and time horizon, which may produce contradictory predictions—for example, estimating a lower cumulative risk of ventilation at 48 hours than at 24 hours. Joint-CE instead models event type and timing together in a single probability distribution, so cumulative event risks cannot decrease over time and event-free probability cannot increase.

Joint-CE was evaluated using two large critical-care databases, MIMIC-IV and eICU-CRD, and compared with a conventional approach using 28 independently trained boosted-tree classifiers. The model maintained comparable or modestly better overall predictive performance while eliminating cross-horizon probability contradictions. It was also transported directly from MIMIC-IV to 55 held-out eICU hospitals without fitting or recalibration on the target database, providing a demanding test of how well the approach transfers across different clinical settings.

Rather than showing that one model can simply achieve a higher accuracy score, this project demonstrates that dynamic ICU predictions can remain internally consistent across both outcomes and time without an evident predictive penalty.

Paper is coming soon….

Thanks for reading my blog !