Hey, I’m Ethan,
an interdisciplinary AI Researcher
With a dual foundation in
computational theory and advanced mathematics,
I research AI systems,
focusing on explainability and interpretability,
along with applications that enhance
quality of life and educational outcomes.
I’ve worked on …
AI-Assisted Grading
AI Validation of Tactile Graphics
Climate Prediction from Satellite Data
Vision Testing in Humans
AI in Robotics Simulations
AI Transfer Learning for Low-Resource Language Speech
Classroom Engagement via Technology
Calibration of Language Models
Deceased-Donor Kidney Allocation
Automated Reasoning for Bounded Pointer Arithmetic
Education
-
Stanford University
M.S. Computer Science
GPA 4.0/4.0, Expected June 2026
-
Stanford University
B.S. Mathematics
GPA 4.0/4.0, 2025
Research and Projects
Vision Response with Kolmogorov-Arnold Networks
Winter 2025 — present
Vision test response correctness probability is a function of font size and visual acuity. Use Kolmogorov-Arnold Networks (KANs) to implicitly learn this probability function from a dataset with known font size and unknown visual acuity. Design a library inspired by pykan to enable more control over learnable activation functions for interpretability and correct learning. Apply library to similar problems in other applications (e.g. hearing response).
Research Mentor: Chris Piech, Assistant Professor of Computer Science, Stanford University.
Research Assistant, Stanford University
Automatic Validation of BANA Guidelines for Tactile Graphics
Summer 2025
Adapting instructional materials for blind and low-vision users requires professional (sighted) tactile graphics designers to convert visual images to tactile graphics, which is nontrivial due to differences between sight and touch. The Braille Association of North America (BANA) defines clarity and consistency guidelines, but no automatic method exists for professionals to check their work against these guidelines. Explored methods for automatic verification, including manual coding at the vector graphics level, use of local AI models, and use of VLM APIs. Project produced a prototype gradio web interface using VLM APIs to analyze proposed tactile graphics for BANA guideline compliance.
Research Mentor: Hariharan Subramonyam, Assistant Professor (Research) of Education and, by courtesy, of Computer Science, Stanford University Graduate School of Education.
Simulation-Based Policy Training for Costly-Data Scenarios
In many RL applications, collecting real-world data can be inherently costly due to reward-independent operating costs (e.g. robotics, human feedback) or real harm from low-reward trajectories (e.g. medicine, robotics). This project implements a non-model-specific simulator, allowing exploration of a less-accurate simulated environment with on-policy training methods without incurring extra data collection costs. Training alternates between the simulated and true environments, with all new true environment data used to improve the simulator over the course of training. CS 224R, Spring 2025
Reinforcement Learning-Based Exam Grading Assistant
Exam questions for classes like CS 109 (introductory probability) cannot be reused, and are time-intensive for humans to grade. This project accelerates rubric-based grading for CS 109 by semi-supervised clustering on semantically similar student responses. A graph-based RL model determines which ungraded student response to ask the human to grade in order to most quickly achieve an accurate clustering with a minimum of human-given grades. CS 234, Winter 2025
Neural Networks and Model Explainability for Graphs
Graph Neural Networks (or GNNs) are Machine Learning models that work with data structured as a graph. Explainability in AI is the science of understanding how and why models give the results they do. CS 224W, Fall 2024
Research Assistant, Stanford University
Allocation of Deceased-Donor Kidneys for Transplant
Summer 2024
Used tools such as pandas to analyze discard rates, congestion, and geography-specific changes in kidney placement before and after a 2021 policy change. Cross-referenced historical data with policy documents and experiences of medical personnel to produce concise, self-contained, and reproducible reports.
Research Mentor: Itai Ashlagi, Professor of Management Science and Engineering, Stanford University.
Combining Contrastive Learning with Layer Utility Analysis and Experimental Multi-Task Finetuning to Improve mini-BERT Performance
Showed which hidden layers of 12-layer BERT best encode required information for each of three tasks (Sentiment Classification, Paraphrase Detection, Semantic Textual Similarity) by attaching and training classification heads of several types to each layer. Overall project goal was to improve BERT performance on the three downstream tasks. Recommended minimum number of pretrained layers for each task was 8, 4, and unclear (but 12 works), respectively. CS 224N, Spring 2024
Calibrating Confidence Measures for BERT for Sentiment Classification
Tested calibration of BERT model for Sentiment Classification (5-point sliding-scale). Evaluated an array of confidence metrics for calibration and correlation against model (in)accuracy. Proposed and tested a novel confidence penalty for sliding-scale linear classification tasks, finding it improves some confidence metrics but not others. CS 281, Spring 2024
ATDS Scores as a Metric for Continued-Pretraining Effectiveness for Donor Languages for ASR
Continued work on a research effort by Nay San evaluating Acoustic Token Distribution Similarity (ATDS) as a metric for predicting how well a “donor” language will improve multilingual ASR for a target low-resource language when the donor language is used for continued pretraining. If effective, the metric would reduce wasted compute usage. Found ATDS computed at middle and final layer extractions corresponds best to actual WER improvements. Found ATDS is robust to domain shift, but cannot be used for cross-domain comparison of candidate donor audio. CS 224S, Spring 2024
Urban Hot Spot Prediction via Supervised Learning
Implemented and tested several families of ML algorithms for predicting urban hot spot locations from satellite images of cities, reducing the need for on-the-ground temperature measurement data. CS 229, Fall 2023
Implementing and Optimizing a Bounded Pointer Arithmetic Decision Procedure
Bounded pointer arithmetic (BPA) is a satisfiability modulo theory (SMT) that models pointer arithmetic in C. Implemented an SMT-LIB2 interface for BPA formulas, a conversion from BPA to equisatisfiable QF_UFLIA formulas in SMT-LIB2, and benchmarks for BPA. CS 257, Fall 2023
Polynomial Convexity of Simple Complex Shapes
Stanford Undergraduate Research in Mathematics
Summer 2023
Researched the polynomial convexity of collections of disjoint simple shapes in complex space. Poster presented at 2023 Stanford Symposium of Undergraduate Research.
Characteristics of glued numerical semigroups and the Kunz cone
San Diego State University Mathematics REU
Summer 2022
Researched the Apery posets of glued numerical semigroups and faces of the Kunz cone containing glued numerical semigroups.
Results presented at the 2023 Joint Mathematics Meetings.
Fibonacci Random Generator and Fourier Analysis
Stanford Undergraduate Research in Mathematics
Summer 2021
Researched the mixing time of the Chung-Diaconis-Graham random process and its generalizations, with particular attention to the Fibonacci Random Generator.
Technical Skills
-
Languages
Python
Wolfram Language
C and C++ -
Python Libraries
PyTorch
PyTorch Geometric
Gradio
google-genai (Google Generative AI API)
Pandas
pykan
TensorFlow -
Tools
LaTeX
Mathematica
Google Cloud
AWS
SVG (specification)
Teaching
Stanford University
CS109
Probability for Computer Scientists
I am a CS109 teaching assistant for Chris Piech and Jerry Cain.
Relevant Coursework
Computer Science
CS 230: Deep Learning
CS 234: Reinforcement Learning
CS 224R: Deep Reinforcement Learning
CS 224N: Natural Language Processing with Deep Learning
CS 224S: Spoken Language Processing
CS 281: Ethics of Artificial Intelligence
CS 362: Research in AI Alignment
CS 229: Machine Learning
CS 224W: Machine Learning with Graphs
CS 257: Intro to Automated Reasoning
CS 265: Randomized Algorithms and Probabilistic Analysis
CS 166: Data Structures
CS 161: Design and Analysis of Algorithms
CS 109: Intro to Probability for Computer Scientists
Mathematics
MATH 151: Intro to Probability Theory
PHIL 151: Metalogic
MATH 108: Intro to Combinatorics and its Applications
MATH 145: Algebraic Geometry
MATH 154: Algebraic Number Theory
MATH 155: Analytic Number Theory
MATH 175: Elementary Functional Analysis
Entrepreneurship
MS&E 178: The Spirit of Entrepreneurship
MS&E 472: Entrepreneurial Thought Leaders Seminar
EDUC 254: Digital Learning Design Workshop
EDUC 295: Entrepreneurship and Innovation in Education Technology
Linguistics
LINGUIST 1: Intro to Linguistics
LINGUIST 167: Languages of the World
PSYCH 1: Intro to Psychology
2 quarters of Italian (1 quarter abroad in Florence, Italy)
4 years of high school Latin, National Latin Exam Gold Medal winner
