About Me

(Last updated 2025-07-12)

Hi there, I'm Luke Merrick, a member of the technical staff at Datology AI.

Below is some of my writing, code, etc.

Writing

2025

• Are We Entering A First-Order Optimizer Renaissance?, musings on recent progress made by Muon and other newly developed algorithms for training large neural networks faster and better

2024

• Arctic-Embed 2.0: Multilingual Retrieval Without Compromise, the technical report that tells the story of Arctic Embed 2.0
• Arctic-Embed: Scalable, Efficient, and Accurate Text Embedding Models, the technical report describing the initial development of the Arctic Embed text embedding models
• Embedding And Clustering Your Data Can Improve Contrastive Pretraining, a writeup from a research side-quest to improve contrastive pretraining in modeling efforts like Artic Embed

2023

• JuLox: What I Learned Building a Lox Interpreter in Julia

2022

• Optimizing utility-scale battery storage dispatch (Medium), published on the Myst.ai blog
• Using volatility forecasts to reduce risk and increase returns (Medium), published on the Myst.ai blog

2021

• The Uncanny Dominance of the Python ML Ecosystem, thoughts on why a loose collection of Python packages stays the perennial popular choice for Machine Learning

2020

• Recursive Least Squares Derivation walks through Recursive Least Squares and derives a batch update equations friendlier to deep learning applications

2019

• The Explanation Game: Explaining Machine Learning Models Using Shapley Values (arXiv), published in CD-MAKE 2019, best paper award
• Confidence intervals for permutation importance (Medium, paywalled) published on Towards Data Science
• Randomized Ablation Feature Importance (arXiv) (permutation feature importance), technical writeup

Open Source

• Arctic Embed training code, an open-source training loop for Arctic-Embed-style text embedding models. Includes custom data loading, loss functions, and an in-depth example.
• Lodis: Local Dictionary Server, a Python-native dictionary server for cross-process communication backed by shared memory from the operating system.
• forecasting_for_utility_scale_battery_optimization.ipynb (GitHub), a notebook walking through forecast-driven optimization of utility-scale battery assets.
• Nearest neighbor vector retrieval using integer quantized embedding vectors.
• JuLox.jl (GitHub), an implementation of the Lox programming language from Crafting Interpreters in Julia. I also have a blog post about JuLox.
• InvestorsExchange.jl (GitHub), a tool to parse trade-level stock data from the IEX exchange in Julia. There was quite a reaction on HackerNews because automatic recapitalization made the post title comically misinterpretable.
• RollingTimeWindows.jl (GitHub), a tool for indexing time-indexed data in fixed-duration periods, even when the number of rows per period is not fixed.

External Links

If you're interested in chatting about ML, quantitative trading, or explainable AI, reach out!

• Email
• GitHub
• LinkedIn
• Google Scholar
• Twitter