Machine Learning Research, Systems & Engineering
Exploring model architectures, building reliable ML platforms, and engineering software with AI coding agents.
Independent research and engineering notes drawn from real implementations, experiments, and lessons learned, including what didn't work.
Featured writing
-
When Does Model Complexity Pay Off?
Lessons from a short-horizon time-series bake-off: sequence models out-ranked gradient boosting by 0.016 to 0.030 AUC, and higher AUC did not reliably translate into better decisions.
-
Building a Point-in-Time Correct ML Feature Store
Point-in-time joins don't guarantee causal features. How we built a feature store with seven enforced safeguards—and uncovered four data leaks that had passed earlier validation.
-
Can AI Coding Agents Build Reliable Software?
Across two engineering projects and nearly 3,200 commits, AI coding agents built software, introduced defects, and implemented safeguards that exposed them. What the failures revealed about engineering reliability, enforceable rules, and human oversight.
-
Why Offline ML Evaluation Fails
Eleven documented failures of look-ahead leakage, selection bias and offline/live mismatch survived conventional validation. How prefix tests, historical-data audits and serving-path replay exposed them—and the engineering safeguards built in response.
-
Numerical Reproducibility in ML Systems
A reference's answer is its arithmetic, not its formula. Comparing seven reimplemented packages with scikit-learn, scipy and pandas, bit-exact where achievable, exposed fused multiply-adds, summation orders and stopping rules the documentation never states—and 38 defects underneath.
-
Crow, a Programming Language
A small general-purpose language that reads like Python, compiles to Rust, runs the same program interpreted or compiled, and builds web applications with no JavaScript. What the simplicity buys, where the speed comes from, what you can build with it today, and how to try it in a minute.
-
Building Agents in Crow
What agents, harnesses, skills and MCP are, built up from a single model call, and how Crow builds each one in its standard library.
-
How Fast Is Crow?
Crow runs one program two ways: interpreted, the moment you start it, or compiled to a native binary through Rust. This article measures both against Rust, Go and CPython on eight programs from the Computer Language Benchmarks Game, on one machine, with every port's output checked byte for byte.
-
An Overview of Information Retrieval Methods
An information retrieval system is built from parts that work at different stages: some plan what to search for, some find candidates, some combine and reorder them, and some turn what was found into an answer.
Independent projects
