Research & Projects
SELECTED WORK · 2022–2026
Research that diagnoses failure—and engineering that ships.
I work across reliable LLM reasoning, information retrieval, computer vision, and applied machine learning. The common thread is evaluation: understanding what works, where it breaks, and what the evidence actually supports.
SciTaRC: Scientific Tabular Reasoning
Co-developed an expert-authored benchmark that tests language understanding, strategic planning, and complex computation over tables from scientific papers. Our study identifies execution compliance—not strategy formation—as a dominant bottleneck.
How Grounded is Wikipedia?
Contributed to a benchmark of approximately 150,000 annotated claims from 1,485 biography articles and evaluated lexical and dense retrieval systems for tracing claims to supporting evidence.
Debugging Reasoning in Text Games
Investigated failure modes and stuck states in interactive fiction using TALES and MindCoder, with emphasis on long-horizon decisions and targeted intervention.
AstroPath Cell-Structure Segmentation
Supported region-level cell segmentation with UNSEG and CelloType and evaluated outputs using DeepCell metrics for the AstroPath initiative.
Robust Active Learning under Distribution Shift
Explored G-optimal design, shift-pessimistic querying, and density-ratio calibration while validating pipeline behavior under distribution shift.
Code-Switched Neural Machine Translation
Built a multilingual translation system with language-aware representations and multi-task learning for code-mixed inputs.
Object Detection & Tracking
Developed an end-to-end gameplay detection and tracking pipeline, reaching 92% detection accuracy at 34 FPS.
Audio-to-Image Generation
Built a variational autoencoder that maps spoken digit audio to handwritten-style images and evaluated outputs with a fine-tuned classifier.
Travelers Modeling Competition
Led an XGBoost-based conversion modeling project on more than 80,000 records. Improved AUC by 13%, reduced RMSE by 21%, and proposed strategies with an expected 8% conversion uplift.