current · ai · concept
Machine Unlearning and Concept Scaffolding
Formalizing and testing how concepts are represented in neural networks, with the goal of unifying diverse phenomena across machine learning
read moreI research and build human-centered AI systems.
MIT senior passionate about AI research, technology & its interface with society.
about
About Eva
My work sits at the intersection of AI, human systems and mathematical abstraction. I study how meaning, structure, and intelligence arise in complex systems, whether in people, language models, or the interaction between the two.
I’m especially interested in the hidden representations that shape behavior: how humans encode identity and culture through language; how machines encode concepts through geometry; and how aligned intelligence can emerge when these worlds meet.
Ultimately, my goal is to design safe, fair and ethical AI systems that enhance rather than detract from human creativity and capacity.
Current interests
- Limitations of LLMs and mind-machine asymmetries
- Information geometry and representation topology
- Multiagent learning, reinforcement learning
- Machine unlearning and concept scaffolding
- AI companions that adapt to linguistic style
- Emotional + cognitive load in human-AI interaction
- AI applications across diverse social and industrial contexts
- Long term second-order effects of AI systems
projects
Selected work
current · ai · prototype
Style-adaptive AI companion
Investigating the limitations of LLM architecture that contribute to poor ability to emulate linguistic style. Early-stage experiments in LLMs that adapt to user language, emotional tone, and cognitive style.
notes / repo (soon)current · ai · research
Optimism in Proximal Policy Approximation (PPO)
Investigating whether optimistic gradient updates to PPO can improve multiagent reinforcement learning (MARL). Starts from theoretical derivation of update rule and evaluates through simple games and an extensive-form imperfect information game.
paper (soon)current · ai · prototype
A Pluralistic LLM Implementation
Implementing a distributional, pluralistic LLM with Retrieval-Augmented Generation (RAG) that reflects the diversity of campus opinion trained on MIT Confessions data
paper (soon)2024 · ai · research
Language Models as Dual Process Reasoners
Co-authored paper introducing “dual prompting,” a novel NLP prompting strategy inspired by Kahneman’s dual-process theory; evaluated on MATH, GSM8K, and MathQA datasets with LLaMA 3.1.
paper2024 · web dev · prototype
World Dictionary
Created full stack web app solution to language revitalization: an online dictionary where anyone can post definitions, but only fluent speakers can verify
websiteexperience
Roles & experience
BS in Computer Science · MIT
Selected coursework: Topics in Multiagent Learning, NLP, Theory of Computation, Algebra 1, AI Decision Making & Society
TPM Intern · Microsoft
Led end-to-end UX redesign of W365 Cloud PC provisioning, including ideation, customer research and prototyping. Culminated in Figma prototype and product spec presented to leadership, to be deployed in Q4 of 2025
TPM Intern · Microsoft
Led customer research on AI and cloud integration to inform feature prioritization and product spec. Performed competitive analysis to drive business decisions on cloud backup solutions. Created product demo with Powerpoint and Clipchamp
Frontend Development Intern · Brittany Gene Design
Developed frontend for the Navajo Nation scholarship application portal. App is first of its kind and will be used as stepping stone for wider digital modernization projects within the Navajo Nation government
writing
Selected writing
ai
What Is the 'Minimal Essence of Experience' for Machine Learning?
A new lens on concept formation, learning and unlearning in AI
read moreproduct
A self-compassionate system for breaking the procrastination cycle
A framework for tackling procrastination at its root
read moreai
The generic style of LLMs
On the architectural limitations of LLMs that contribute to their generic linguistic style and poor ability to match user language
read morecontact
Let’s talk
I’m open to roles, collaborations, and conversations. If something here overlaps with what you’re building, I’d love to hear from you.