Matthew Heng
Diamond Bar, California.
Hello, I'm Matthew, a junior at Diamond Bar High School. Most of my time outside class goes to research — currently at USC's Autonomous Networks Research Group, on how small a language model can get before it stops writing coherently. I've also been learning quantum software and bioinformatics, mostly by getting things wrong first. More in /research, /projects, /school, and /aboutme.
Research
-
Transformer language models from 78K to 989K parameters, trained on TinyStories and targeted at an ESP32. An offset power law fits the width sweep at R² = 0.998, with no saturation across the whole sub-million-parameter range. Beats the released TinyStories 1M checkpoint on bits per byte with 3.8× fewer parameters. USC ANRG, under Prof. Bhaskar Krishnamachari.
-
Do Large Language Models Deceive Under Pressure?
First author. Constructing simulated high-stakes conditions to test frontier models for deceptive decision making.
-
NASA GeneLab for High Schools
Twelve weeks of online space-biology and bioinformatics training run by NASA Ames' Space Biosciences Division, at roughly 20 hours a week. Omics-based analysis of spaceflight transcriptomics from NASA's Open Science Data Repository, worked through in JupyterLab notebooks. Completed the core course, which carried eligibility for the program's optional Capstone Project.
-
MIT Beaver Works, Quantum Software
Four weeks on quantum software with MITRE, admitted after prerequisite coursework in Python, systems, and the background mathematics. Built the foundation from linear algebra — complex vector spaces, unitary operators, eigendecomposition — through bra–ket notation and tensor products to superposition and multi-qubit measurement, then implemented Deutsch–Jozsa in Q# using CZ phase oracles and phase kickback. Final project: a variational quantum eigensolver for the transverse-field Ising model, written in Q# and benchmarked against exact diagonalization across ansatz depths and boundary conditions.
Projects
-
Co-founder. A nonprofit putting model rocketry kits and aerospace curriculum in front of students without access to them. I built the site that handles registration and content delivery.
-
FireShield
A Swift computer vision app that separates ordinary heat sources from genuine fire hazards, optimized for on-device inference. Presidential AI Challenge submission.
-
Teaching Python, Java, and JavaScript to twenty-odd elementary and middle school students, through curricula built around logic and debugging rather than syntax.
School
Awards
-
Gold Medalist, Calico Hackathon at UC Berkeley
31st of roughly 500 participants; gold awarded to the top 44.
-
National Qualifier, FBLA Coding and Programming
Advanced from state on programming logic, algorithm design, and software development.
-
State Qualifier, FBLA Cyber Security
Qualified on network security, encryption, and digital forensics.
About me
Outside of research I play the oboe, and I'm usually the one saying yes to plans made an hour in advance. Lately I've been playing a lot of poker with friends. Mostly I like learning things I'm bad at, which is the only through-line I can find across microcontrollers, qubits, and gene expression.