Experience
Software Engineer Intern
PayPal
- ▹Built AI-powered tools that resolve discrepancies between PayPal's internal and external-facing API error codes, helping internal engineers debug faster and external partners integrate more reliably
- ▹Deployed as a company-wide MCP server containerized with Docker, already used to fix production bugs
Graduate Researcher
Software Systems Lab, Columbia University
- ▹Developed a framework for fine-tuning vision-language models for classification tasks with high intra-class diversity
- ▹Built fault-tolerant training infrastructure on a GPU cluster with automated checkpointing and job recovery
Graduate Researcher
Zemel Group, Columbia University
- ▹Built a research framework to evaluate how well LLMs can write formal mathematical proofs under constrained inference-time compute, enabling fair comparison across different reasoning strategies
- ▹Benchmarked open 7-8B parameter LLMs on the audited miniF2F proof dataset, using the Lean 4 assistant to verify correctness and a custom metric to measure proof success rate under a fixed compute budget
Undergraduate Researcher
Zuckerman Mind Brain and Behavior Institute, Columbia University
- ▹Evaluated Temporal Convolutional Network models for classifying mouse behavior, analyzing multi-modal time-series data (wheel velocity, paw locations) to validate predictions
Education
Sep 2025 – May 2026
Master of Science in Computer Science
Columbia University
Fu Foundation School of Engineering and Applied Science
- ▹Researcher in 2 CS labs
- ▹Focus: Machine Learning, Computer Vision, Automated Theorem Proving
Sep 2021 – May 2025
Bachelor of Arts in Computer Science
Columbia University
Columbia College
- ▹Dean's List
- ▹Relevant coursework: Databases, High-Performance Computing, Machine Learning, Computer Vision, Natural Language Processing, Robotics, Analysis of Algorithms, Systems Programming, Computer Systems, Probability, Statistics
Projects
Bipedal Locomotion
Lightweight MuJoCo + PyTorch framework for training and evaluating bipedal locomotion policies with Proximal Policy Optimization, using GAE-Lambda advantage estimation and a clipped surrogate objective. Includes Actor-Critic networks, inverse kinematics, and support for both 2D and 3D robot morphologies.
Granite Speech in FMS
An 8-week Columbia HPML research collaboration with IBM Research to port IBM's Granite Speech 3.3 8B model — a Conformer encoder, Q-Former projector, and LLM decoder speech-to-text architecture — out of Hugging Face Transformers and into IBM's Foundation Model Stack (FMS), a native PyTorch framework built for compiled, production-grade inference, so it runs end-to-end under torch.compile.
Action Segmentation in Mice
Diagnostic and behavioral-fingerprinting pipeline built around a Temporal Convolutional Network model, evaluating its classification of mouse behavior from multi-modal time-series data across 57 research sessions.
Life Expectancy Analysis
Multiple linear regression analysis in R examining which socioeconomic and political factors most strongly predict national life expectancy, with model selection cross-checked across three criteria and validated two independent ways.
LLM Alignment Preservation
A research paper on the "alignment tax" — the tendency for a safety-aligned LLM to lose that alignment as it's fine-tuned on unrelated, entirely benign capability tasks. Reframes the problem as catastrophic forgetting and runs the first evaluation of state-of-the-art continual learning methods across a realistic 7-task sequential fine-tuning chain, instead of the two-stage settings prior work relied on.
State Space Model POS Tagger
Part-of-speech tagger implemented in MATLAB using a structured state-space model with a bilinear-discretized state transition matrix and manually-derived backpropagation.
CKY Parser
Cocke-Kasami-Younger chart parser for context-free grammars, implementing probabilistic (Viterbi) parsing with a full Parseval evaluation harness.
Wikipedia Article Clustering
Live-scraped Wikipedia articles, embedded with Sentence-BERT, and clustered with UMAP + K-Means to surface topic structure — visualized as an interactive scatter plot.
Neural Network Dependency Parser
Transition-based dependency parser using a neural network architecture implemented in PyTorch, trained on Penn Treebank annotations.
TDNN-Conformer ASR
Novel speech recognition architecture combining Time Delay Neural Networks with the Conformer model. Implemented and trained a 12.5M parameter model using PyTorch and WeNet on the LibriSpeech dataset.
Hierarchical Image Classifier
Bird, Dog, or Reptile? — a two-headed ResNet-50 that predicts both a coarse category and a fine-grained sub-class simultaneously, with confidence-threshold detection for categories never seen in training. Built with two Columbia classmates.
Trigram Language Model
Statistical language model using trigram probabilities with linear interpolation smoothing for text generation, perplexity evaluation, and TOEFL essay proficiency classification.
Wonderful Workouts
Weight training education platform with a hand-built HTML5 drag-and-drop routine builder, muscle-group coverage validation, a knowledge-check quiz, and PDF export.
Home Cooking Platform
Full-stack Flask + PostgreSQL web application with a 10-table relational schema and dynamic filtering that recommends recipes based on available ingredients, dietary restrictions, cuisine, and complexity — matching all selected constraints at once.
Bipedal Locomotion
Lightweight MuJoCo + PyTorch framework for training and evaluating bipedal locomotion policies with Proximal Policy Optimization, using GAE-Lambda advantage estimation and a clipped surrogate objective. Includes Actor-Critic networks, inverse kinematics, and support for both 2D and 3D robot morphologies.
Granite Speech in FMS
An 8-week Columbia HPML research collaboration with IBM Research to port IBM's Granite Speech 3.3 8B model — a Conformer encoder, Q-Former projector, and LLM decoder speech-to-text architecture — out of Hugging Face Transformers and into IBM's Foundation Model Stack (FMS), a native PyTorch framework built for compiled, production-grade inference, so it runs end-to-end under torch.compile.
Action Segmentation in Mice
Diagnostic and behavioral-fingerprinting pipeline built around a Temporal Convolutional Network model, evaluating its classification of mouse behavior from multi-modal time-series data across 57 research sessions.
Life Expectancy Analysis
Multiple linear regression analysis in R examining which socioeconomic and political factors most strongly predict national life expectancy, with model selection cross-checked across three criteria and validated two independent ways.
LLM Alignment Preservation
A research paper on the "alignment tax" — the tendency for a safety-aligned LLM to lose that alignment as it's fine-tuned on unrelated, entirely benign capability tasks. Reframes the problem as catastrophic forgetting and runs the first evaluation of state-of-the-art continual learning methods across a realistic 7-task sequential fine-tuning chain, instead of the two-stage settings prior work relied on.
State Space Model POS Tagger
Part-of-speech tagger implemented in MATLAB using a structured state-space model with a bilinear-discretized state transition matrix and manually-derived backpropagation.
CKY Parser
Cocke-Kasami-Younger chart parser for context-free grammars, implementing probabilistic (Viterbi) parsing with a full Parseval evaluation harness.
Wikipedia Article Clustering
Live-scraped Wikipedia articles, embedded with Sentence-BERT, and clustered with UMAP + K-Means to surface topic structure — visualized as an interactive scatter plot.
Neural Network Dependency Parser
Transition-based dependency parser using a neural network architecture implemented in PyTorch, trained on Penn Treebank annotations.
TDNN-Conformer ASR
Novel speech recognition architecture combining Time Delay Neural Networks with the Conformer model. Implemented and trained a 12.5M parameter model using PyTorch and WeNet on the LibriSpeech dataset.
Hierarchical Image Classifier
Bird, Dog, or Reptile? — a two-headed ResNet-50 that predicts both a coarse category and a fine-grained sub-class simultaneously, with confidence-threshold detection for categories never seen in training. Built with two Columbia classmates.
Trigram Language Model
Statistical language model using trigram probabilities with linear interpolation smoothing for text generation, perplexity evaluation, and TOEFL essay proficiency classification.
Wonderful Workouts
Weight training education platform with a hand-built HTML5 drag-and-drop routine builder, muscle-group coverage validation, a knowledge-check quiz, and PDF export.
Home Cooking Platform
Full-stack Flask + PostgreSQL web application with a 10-table relational schema and dynamic filtering that recommends recipes based on available ingredients, dietary restrictions, cuisine, and complexity — matching all selected constraints at once.
drag to scroll