Michael Yao
hi! i'm michael
my goal is to understand how learning systems work, and to use them to do science.
right now i'm a machine learning engineering intern at alkera ai (yc s26), where i train sql agents with reinforcement learning. on the research side, i work with bilin zhuang at harvey mudd on water and the mean-field theory of deep networks, and with miquel duran-frigola at ersilia on tabular foundation models for molecules. i also like reading what training changes inside a model (interpretability!).
i love reading detective stories, a big fan of Shirai Tomoyuki
INTERESTSLLM post-training (RL) · Tool-use agents · AI for science · Mechanistic interpretability
RESEARCH EXPERIENCE
Harvey Mudd CollegeOct 2025 – Present
Undergraduate Researcher · Advisor: Bilin Zhuang · Claremont, CA
Unsupervised Machine Learning on Water Structure
Likelihood-ratio abstention · Molecular Dynamics analysis
Mean-Field Theory of Deep Neural Networks
Mean-field theory · Order-to-chaos boundary
Ersilia Open Source InitiativeApr 2026 – Present
ML Researcher · Advisor: Miquel Duran-Frigola · Remote
Feature Selection for Tabular Foundation Models
TabPFN in-context learning · Molecular feature selection
INDUSTRY EXPERIENCE
Alkera AI (YC S26)May 2026 – Present
Machine Learning Engineering Intern · San Francisco, CA
Training AI agents to fix data pipelines, rewarded only when their code gets the right answer
I built Dataline, the RL environment behind it: a data engineer's whole working computer (warehouse, dbt, scheduler, BI dashboards, tickets, chat), rebuilt from a seed with something broken. The agent gets a terminal and is graded only on what it leaves behind; lying about its work or crossing a permission line is the only way to score below zero.
Fresenius KabiAug 2026 – Sep 2026
R&D Data & Quality Systems Intern · Wuxi, China
Automating quality-compliance checks across hundreds of R&D documents
PROJECTS
NULLSTEP ↗Telling real changes inside a fine-tuned model apart from measurement noise
LANGALPHA ↗AI research agents for finance that show their reasoning and cite their sources
SPEQTRO ↗An AI agent that reads a molecule's spectra and works out its structure
PUBLICATIONS
[1]Y. Yao, M. Duran-Frigola. Feature Selection for Tabular Foundation Models: Exploring Feature Utility in Low-Data Molecular Prediction. arXiv:2609.30789, 2026.
[2]Y. Yao, D. Sanghi, A. D. Chokshi, et al. Is Water a Mixture: Unsupervised Likelihood-Ratio Classification Reveals Two Water States That Merge Upon Increasing Temperature. Under review, PNAS, 2026.
[3]Y. Yao, C. Blodin, et al. A Null-Step Control for Sparse-Autoencoder Feature Drift Across Fine-Tuning Checkpoints. Manuscript, 2026.
EDUCATION
Harvey Mudd CollegeExpected May 2028
B.S. — Joint Computer Science & Mathematics · Claremont, CA
GPA 3.90 · Statistical Mechanics · Probability · Statistics · Differential Equations · Graph Theory
Shanghai Qibao Dwight High School2022 – 2025
IB Diploma · Shanghai, China
Score 44 / 45
PROGRAMS
Break Through Tech AIMay 2026 – Present
Cornell Tech
Summer Science Program (SSP)Jun 2024 – Aug 2024
Southwestern Oklahoma State University · Weatherford, OK
*** END OF PAGE ***