Machine learning researcher · Philadelphia, PA
I build learning systems that remember, adapt, and reason—toward a future where automation creates more room for people to collaborate, create, and live with dignity.
I am an M.S.E. student in Systems Engineering at the University of Pennsylvania, working with the Ale Lab and Huang Lab. My research explores self-evolving memory, continual learning, and useful AI systems that can keep improving in the real world.
Research focus
Self-evolving memory
Continual learning
Human-centered automation
Why I build
I do not expect technology to settle every human question. I do hope to build tools that make more generous answers possible.
01 · THE PROMISE
Its early promise was a more connected world—a “global village” and a better twenty-first century. That future now feels both closer and more fragmented. I still believe computing can help us build something fairer, but I no longer think fairness is simply an optimization objective. Every allocation encodes choices, and even the best-aligned solution can leave someone out.
02 · AUTOMATION
I keep returning to automation: agents that can check their own work, serve people continuously, and learn without erasing what came before. When useful capabilities and essential resources become more available, fewer human problems need to be framed as contests over scarcity.
03 · HOW I WORK
I am drawn to win–win systems—progress that does not require someone else to lose. Code Forest and PathCAR are early, technical steps toward that larger idea: learning systems that evolve, retain knowledge, and stay useful in the real world.
Selected research
Current work spans code reasoning, pathology foundation models, and efficient fine-tuning. Each visual summarizes the project’s core mechanism.
A hierarchical external memory for LLM-based code reasoning.
Code Forest turns solved programming problems into reusable functions and atomic modules. An automated pipeline combines LLM distillation, AST dependency analysis, semantic retrieval, and function merging to support context-efficient code synthesis.
A 41-dataset benchmark and PathCAR, a prompt-conditioned adapter-routing direction.
I benchmark LiteFM, BioMedCLIP, and CONCH across full and parameter-efficient fine-tuning, replay, knowledge distillation, EWC, and LoRA-based continual-learning methods. Weight-space analysis reveals adapter interference and motivates routing specialized adapters by task context.
A practical comparison of LoRA variants and their quality–efficiency trade-offs.
We evaluated LoRA, LoRA-FA, AdaLoRA, and VeRA on Llama-3-8B, with an automated LLM-as-a-judge pipeline measuring generation quality alongside memory and latency on Delta HPC.
Background
2026—present
Huang Lab · University of Pennsylvania
2025—present
Ale Lab · University of Pennsylvania · Advisor: Prof. Osbert Bastani
2024
ECCOM · Built tabular ML pipelines and a 90%+ accuracy employee-turnover model.
2023—2025
University of Rochester · Programming, computational statistics, and computer models.
Education
M.S.E., Systems Engineering · Expected 2027
Previously
B.S., Computer Science & Data Science · 2024