A few things I’ve made
Tools for understanding how models think—and knowing whether they work.
Featured project · Evaluation design
Beyond “seems to work.”
The Hitchhiker’s Guide to Evals. A practical guide to building evaluations that humans and AI can actually use.
Explore the guideGood evals keep getting better.
What good looks like
Start with production traces and define what a successful response should do.
Choose a stage · adapted from the guide’s improvement loop
Where do frontier models share our very human biases?
Lowest bias magnitude score
on the final core leaderboard.
15 core biases. 11 models. A benchmark that ran until the newest generation saturated it.
View benchmarkTesting ideas against the clock. Faster experiments, better questions.
The race to converge.
Bits per byte · lower is better
124M · 1× H200 · March 17, 2026
NCA scalar-fix sweep · no configuration beat baseline.
Research agents explored pretraining ideas and ran experiments. This sweep tested whether preserving learned scalars improved transfer from neural cellular automata.
Read the experimentThe path so far
From understanding people to building and evaluating AI.
Data Scientist
Building evaluations for an internal AI agent harness, with a focus on reliability, latency, and how well agents actually perform.
Currently building evals for one of Meta's internal AI Agent Harnesses on the Sales AI Evals team. Before that: designing, conducting, and interpreting peer-reviewed experiments across multiple verticals, driving roadmap changes such as increased GenAI automation features for the Small Business Group sales teams.
Founder, Alignment Engineer
Founded and shipped three AI-native apps for well-being, personal finance, and fitness. Wound down in 2026.
Founded to impact the human experience through AI-native solutions. Three iOS apps launched: Waves AI (an AI-native well-being alternative to Headspace), LOOT Finance (a Mint and Monarch alternative), and Swole (an agentic, research-based approach to educating lifters while strength-building). Wound down in 2026 when all three apps were deprecated.
Data Scientist
Built measurement tools and analytics infrastructure, reducing job failures by 33% across six queues.
Built internal tooling and scaled infrastructure for analytics and operations teams across global offices.
Sr. Business Analyst / Data Scientist
Built analytics from the ground up for 1,000+ learning centers, generating an approximately 1% lift in franchisor revenue.
Built the analytics practice from scratch, taking the company from bloated Excel spreadsheets to self-serve tooling for the franchisor, franchisees, and employees.
Researcher / Data Informatics Analyst
Made medication adherence and workforce data useful through dashboards and privacy-preserving research.
Developed data visualizations, maintained pipelines, and provided ad hoc research at this healthcare organization with 3,500+ employees.
2020 — 2021
M.S. Applied Analytics
Machine learning, predictive analytics, big data. Graziadio Student Advisory Board Chairman.
2017 — 2020
B.A. Psychology
Statistics, human behavior, experimental design.
Following the questions
Smaller builds, open questions, and tools from along the way.
Published package
Testing whether small models change their behavior when they think they’re being evaluated.
From scratch
Building a GPT in PyTorch to understand every piece, from causal attention to text generation.
Open source
Safety evaluations and crisis-response testing originally built for Waves AI, now retired.
Unfinished prototype
An exploration of linguistic relativity: does the language of a prompt change how a model reasons?
Developer tools
The commands, skills, and hooks I use for everyday development and code review.