Most machine learning projects fail on framing, not on model quality. Five checks to run before you open a notebook: name the decision, price the two kinds of error, audit the target variable, simulate the deployment, and write down how the project fails while the metrics look fine.
WordPress 7.0’s AI updates are heading into Beta 2. This post covers the pivot away from the Abilities API, the new Connector Screen, and the dependency issues in the AI Experiments plugin.
A 420-core Databricks cluster spent 10 hours on 18 partitions because the data was skewed. Here is how salting and liquid clustering fixed the distribution, and what each one is actually good for.
Context engineering means filling an AI model’s context window with your own domain knowledge, using structured graphs, deterministic tools, and persistent memory instead of simple RAG.
Senior data scientist skills are often mistaken for mastery of algorithms. The real gap between junior and senior practitioners is judgment: knowing when to pause, how to frame the problem, and what actually moves the business.
February was short but it made a few things obvious: solo work stalls ML projects, code is not documentation, and an MLOps stack that ignores the deployment environment never ships. Notes from a month of ML and backend work.
Most feature pipelines are just Python scripts and CSV files until they break in production. This post covers using Feast for feature management and Ray for distributed compute to fix training-serving skew and cut latency.
Agentic RAG setups waste time and money answering the same questions repeatedly. This post covers a two-tier semantic and retrieval cache, plus the validation logic needed to keep it from serving stale answers.
Most AI features never leave the notebook. Notes on the production data architecture that gets them into a live site: separated layers, database integrity over raw insert speed, and the state and permission problems agents bring with them.
Sharpness-Aware Minimization finds flatter minima instead of chasing zero training loss, though it doubles training cost and can quietly break your BatchNorm stats if you’re not careful.