Projects

Co-authored research showing that jointly-trained inverse dynamics models produce semantically aligned action embeddings across different game environments. Demonstrated alignment across 7 Atari games, 2 racing simulators, and 2 first-person games (Minecraft VPT and CS:GO), achieving 85.2% accuracy on movement prediction across environments without game-specific labels. Identified a "calibration gap" as the primary failure mode for camera direction prediction and proposed solutions.

Proposed a framework that reduces LLM hallucinations by dynamically selecting from multiple learned correction vectors per prompt, improving factual accuracy by +2.4pp where prior methods gained only +0.3pp. Demonstrated that the framework automatically discovers distinct types of hallucinations without manual labeling, evaluated across 10,615 items spanning 8 knowledge domains.

Built an inference gateway that dynamically routes LLM prompts to right-sized GPU hardware across 4 tiers, reducing energy consumption by 75% on simple queries while reserving high-end GPUs for complex tasks.

Trained a reinforcement learning agent to automatically manage a portfolio across 10 asset classes, achieving 15.76% annualized returns with a 1.44 Sharpe ratio, outperforming traditional strategies by 38-54%.

Built a hands-free AR system integrating AR glasses, a smartphone, and a Raspberry Pi robot, enabling gesture and voice control with live camera feedback — all processed on-device with no cloud dependencies.

Built a VR healthcare simulation on the Apple Vision Pro, enabling CPR and first aid training and pioneering SharePlay integration for collaborative learning with iPhone users. Showcased to The Venture Reality Fund, where the concept was acquired and carried forward.

Built a Bayesian court case prediction framework that achieved 73% top-3 accuracy across 11 possible Supreme Court case outcomes by leveraging observable case attributes. Featured as a winner and future example project in Stanford’s CS109 course (selected from 500 students).

Built an iOS app that deciphers facial expressions to recommend mood‑matching songs, integrating SwiftUI, UIKit, and SwiftData with Google Cloud Vision and Spotify’s Web API. Processed 500+ test images across multiple emotions to validate recommendations, improving user‑song mood alignment accuracy by 82%.

Engineered AI‑based models for medical MRI image reconstruction with PyTorch, cutting required scan times by 55–70% while preserving high‑quality images, and co‑authoring a research paper on the work. Enhanced image quality beyond baseline deep learning models, reducing reconstruction error by ~60% and significantly improving detail preservation.

Engineered an AI agent with OpenAI’s Agents SDK to convert handwritten math to LaTeX with 91% accuracy. Featured as an AgentOps hackathon winner, with board member recommendation for commercial launch.

Designed a music recommendation system using SVD and PCA to analyze over 60 audio features from 7,000 Spotify songs, uncovering patterns in sound beyond traditional metadata. Implemented feature projection to identify similarities across 35 combinations of musical features.