Neha Balamurugan

Neha Balamurugan

Hi, I'm Neha 👋🏾 [NAY-ha] (she/her/hers)

I build and study intelligent systems that help people equitably reason, learn, and create. Here, I share the work and ideas that keep me curious!

Things I care about:

  • Human AI Alignment
  • Educational Equity
  • Femtech
  • Science Integrity

Timeline

  • 2024-2026

    Master's in Computer Science (AI)

    Stanford University

    Causality and Cognition Lab

    Developing benchmarks to study visual social reasoning in humans and AI

    Accel Leadership Program

    1 of 16 students selected for Stanford's flagship entrepreneurship program

    Emerson Collective Venture

    Exploring the software for hardware market

  • 2021-2024

    Bachelor's in Computer Science

    Arizona State University

    🏆 CS Rank #1
    Dreamscape Learn VR Education

    Led 10-member team under $500M immersive learning initiative, built multiplayer VR classrooms

    HCI Studio Berlin

    Designed AR heads-up displays with HoloLens and CARLA simulations to improve driver safety

    Hackathons

    Won multiple hackathons across education, fintech, and health tech

Projects

What I Have Built

  • Spot The Ball

    Under Prof. Tobias Gerstenberg, developed a benchmark to measure how humans and AI infer hidden information in sports scenes. Designed large-scale experiments and evaluated frontier vision-language models to study human–AI reasoning gaps.

    ArXiv
    CogSci 2026 Poster
    • Vision-Language Models
    • Cognitive Science
    • Python
    • Stable Diffusion

  • buff: AI Assistant for Scientific Research

    Undergraduate thesis project developing an AI framework to accelerate scientific discovery by combining large language models, citation graph analysis, and modular retrieval pipelines. Designed to support literature synthesis, knowledge grounding, and experiment planning for researchers across domains.

    Barrett Thesis Symposium 2024
    • LLMs
    • Knowledge Retrieval
    • Research Automation
    • Python

  • Dreamscape Learn VR Education

    Led a 10-member team under the $500M Dreamscape Learn initiative to create multiplayer VR classrooms. Built interactive visualizations for math and climate data, deployed to 1,500+ students across ASU.

    🏆 GCSP Research Award 5x
    ILRN 2024 Conference
    ASU Fulton Forge Conference Posters 2022, 2023, 2024
    • Unity
    • C#
    • VR
    • Data Visualization

  • AR Driver Safety System

    Under the DAAD RISE Fellowship, designed AR heads-up display systems to reduce driver distraction, integrating CARLA simulations, eye-tracking, and real-world HoloLens experiments. Worked with Prof. Thomas Kosch at Humboldt University of Berlin.

    🏆 Barrett Hall of Fame Award
    DAAD RISE Conference Talk 2023
    • AR
    • HoloLens
    • Python
    • CARLA Simulator

  • Fluid Dynamics Visualization in VR

    Supported by the NSF Research Experience for Undergraduates program to build interactive VR systems for exploring large-scale fluid simulations, improving real-time rendering speed by 45%.

    • VR
    • Unity
    • VFX Graph
    • GPU Optimization

  • Recall: Wearable Conversation Recorder

    CalHacks '25

    A wearable built on Raspberry Pi that auto-detects conversations via voice cues, generates FaceNet embeddings, and records local audio. Transcripts are summarized with Gemini and stored in Supabase for natural-language search in a Next.js app.

    • Raspberry Pi
    • FaceNet
    • Gemini
    • Next.js
    • Supabase

  • YouCare: Health App For Women's Wellness Forecasting

    TreeHacks '23

    Web application to enhance women's health awareness through personalized symptom analysis, incorporating Apple HealthKit API and period tracking apps like Flo. Built with Swift in XCode and integrated with the You.com AI search platform.

    🏆 Best Use of Open Platform Award
    • Swift
    • XCode
    • Apple HealthKit
    • You.com AI

  • SusProduce: AI Vision For Post-Expiry Food Usability Assessment

    HackMIT '23

    ML image detection and analysis tool utilizing the Keras library to discern the usability of food past its expiration date. Developed an Angular front-end application aimed at minimizing food waste through intelligent classification.

    🏆 Best Solution for CPG
    • Keras
    • ML
    • Angular
    • Image Classification

  • CareUp: Disability Caretaker Assistant

    Hacks for Humanity '23

    Created an AI-powered app that simplifies caregiving by finding, compiling, and ranking resources and programs for elderly or disordered family members, and assists caretakers in completing necessary forms.

    🏆 1st Place
    • AI
    • Web App
    • Healthcare

  • Althea: AI Research Agent

    TreeHacks '24

    An AI-powered research assistant that helps researchers conduct literature reviews and identify research gaps. Fine-tuned on ~500 biochemistry papers with citation network integration and semantic chunking for efficient information retrieval.

    • Python
    • Reflex
    • LangChain
    • OpenAI
    • MongoDB
    • Pinecone
    • LLMs

  • Navigating the Lunar Surface under Low Visibility Conditions

    Decision Making Under Uncertainty · Fall 2024

    Developed a grid-world simulation to model lunar rover navigation under uncertainty, comparing QMDP-based decision-making with fully observable and random agents. Demonstrated that QMDP effectively balances exploration and planning under partial observability, achieving up to 100% success in dense environments.

    • RL
    • POMDP
    • Autonomous Navigation
    • Robotics

  • Investigating the Temporal Sense of Large Language Models

    NLP · Winter 2025

    Studied how large language models represent and reason about time by modifying the TOT-Semantic benchmark to include natural language and linguistic noise. Evaluated Gemini, Mistral, and LLaMA on diverse graph structures, revealing that models rely more on pattern recognition than true temporal reasoning.

    • LLMs
    • Temporal Reasoning
    • Benchmarking
    • Natural Language Understanding

  • Volumetric Rendering and Visualization in a Custom Ray Tracer

    Computer Graphics · Winter 2025

    Implemented volumetric rendering in a physically based ray tracer, simulating absorption, in-scattering, and phase function effects to visualize light transport in fog, smoke, and translucent media. Built interactive UI controls for density and scattering parameters, added Perlin-noise–based procedural volumes, and designed visualizations of ray marching to make rendering concepts intuitive.

    • Computer Graphics
    • Ray Tracing
    • Volumetric Rendering
    • Visualization

  • Improving Small Language Models via Test-Time Prompt Compression and Retrieval

    Deep RL · Spring 2025

    Investigated how retrieval and compression strategies affect small language models at inference time. Proposed a hierarchical framework using Gemma 12B for prompt planning and Qwen2.5-0.5B for generation, showing that concise prompt compression improves reasoning quality without additional training.

    • LLMs
    • RAG
    • Prompt Compression
    • RL

  • DSPy: ContextSeeker

    Visual Computing Systems · Spring 2025

    Developed a new DSPy module enabling language models to ask clarifying follow-up questions when user prompts lack key context. The ContextSeeker framework introduces trainable components for question generation, stopping criteria, and human-in-the-loop optimization, improving reasoning accuracy on underspecified tasks like Fermi problems.

    • LLMs
    • Human-in-the-Loop
    • DSPy
    • Interactive AI
    • Prompt Optimization

  • Estimating Fuel Efficiency of Aircrafts using GNNs

    ML with graphs · Fall 2025

    Built a Graph Neural Network to predict aircraft fuel consumption using real-world trajectory data from the 2025 Performance Review Commission (PRC) Data Challenge. Modeled flights as spatiotemporal graphs capturing dependencies between consecutive flight segments and airports, improving fuel burn estimation over traditional independent models.

    • GNN
    • Spatiotemporal Modeling
    • Aviation

  • Learning Primal Heuristics for Neural Network Verification

    AI for Alg/Optimization · Fall 2025

    Extends differentiable integer linear optimization (DiffILO) to train heuristic policies for neural network verification (NNV) without supervision. Introduces a differentiable, unsupervised framework that learns MILP heuristics using probabilistic relaxations and CVaR-based feasibility measures, enabling gradient-based optimization of verification problems.

    • Neural Network Verification
    • MILP
    • Optimization
    • Differentiable Programming

Activities

What I'm Involved In

  • Sep-Dec 2025

    CS238· Graduate TA

    Supported Stanford’s graduate course Decision Making Under Uncertainty with Professor Mykel Kochenderfer, helping students navigate topics such as Markov Decision Processes, reinforcement learning, and POMDPs through lectures, office hours, and project mentorship.

  • Heidelberg Laureate Forum - Image 1 of 3
    September 2025

    Heidelberg Laureate Forum· Scholar

    One of 200 researchers in CS/Math selected to attend the conference. Met laureates like Vint Cerf, Richard Sutton, and David Silver.

  • Second Order Podcast
    2025 — Present

    Second Order Podcast· Co-host

    A podcast unpacking the mental models behind impactful entrepreneurship and leadership with conversations featuring guests like Elad Gil.

  • Accel Leadership Program
    2025

    Accel Leadership Program· Fellow

    1 of 16 students selected to participate in Stanford's flagship entrepreneurship leadership program. And now TAing ALP 2026!

Thoughts

What I'm Thinking About

  • Dec 2025

    Software for Hardware Market

    Why software for hardware is the new industrial revolution, and companies to watch in the space.