I built a novel LLM agent that turns raw reliability and latency metrics from Microsoft’s advertising platform into clear, real-time insights — cutting on-call manual work by 22%. I integrated the agent into Grafana's backend API, then led a full migration from PowerBI to Grafana dashboards, giving teams scalable, live data visualizations and transforming how engineers monitor the ads system.

A conversational AI tool for home chefs looking to clean out their fridge. Tell it what ingredients and which dietary preferences you have, and it will generate a recipe for you! Compared to OpenAI's GPT-4o, it generates recipes that are 40% more likely to be rated as 'good' by users. Project for MIT's 6.8611 Quantitiative Methods for Natural Language Processing.

I built a multimodal ensemble model that detects whether cardiovascular wearables are being worn — achieving 99% accuracy. Working with billions of rows of complex accelerometer and ECG data, I optimized large-scale data pipelines, cutting training time by 14 hours (62%). Finally, I engineered and deployed the model into a cloud-backed mobile app, bringing advanced health monitoring directly to patients’ devices.

I built backend services, RESTful APIs, and modern UIs for a funding workflow platform used by asset managers. I automated balance sheet generation, compliance checks, and interdepartmental communication with a new electronic task management system. I also developed a collateral valuation solution (using Java, Spring Boot, React, SQL, Python, SingleStore, Flask) that reduced admin burden by 36%, saving both the firm and its clients significant time.

At the MIT Media Lab, I engineered a generative AI model that emulates Marvin Minsky’s improvisational style using GANs. I developed a digital pipeline for time-series audio data, enabling new instruments to be embedded so musicians can “improvise” alongside Minsky. Leading the signal processing team, I also created a clustering algorithm with 85% accuracy for classifying audio similarity, bridging AI, creativity, and music.

As an undergraduate researcher at MIT CSAIL, I built deep learning infrastructure to reduce demographic, social, and ethical bias in ureteral stone clinical risk scores — achieving 89% accuracy. I contributed across the full development cycle, from testing and debugging to API documentation. I also queried and processed 10,000+ rows of raw patient data from the MIMIC-IV database, turning complex medical records into a usable dataset for machine learning.

At Aster, a healthcare startup, I built machine learning systems for personalized drug recommendations and used NLP to automate precharting of medical information. I integrated front-end webpages for ambient clinical documentation with back-end ML models, creating a seamless physician workflow.

Developed a multimodal deep learning model that combines patient record data with ultrasound kidney images to predict acute kidney injury (AKI), achieving 98% accuracy. This work was presented at the International Conference of Artificial Intelligence in Medicine Applications (London, Sept. 2021), highlighting its potential impact on clinical decision support and early diagnosis.