Your First Model Should Be Embarrassing
A majority-class guess and a logistic regression take seconds to fit. Without them you can't tell which part of the impressive model is doing the work.
Related course: ML101 — Introduction to Machine Learning
Articles
Longer pieces on the judgment calls behind applied data science and AI work.
Your Agent Has Observability. It Doesn't Have Evals.
89% of teams instrumented their agents. 52% check whether the agents are right. The missing half is the half that catches the failure.
Related course: AI401 — AI Applications with LLMs
Text-to-SQL Accuracy Isn't a Model Problem. It's a Schema Problem.
Three studies reported three very different accuracy numbers for the same task. None of them is wrong — they're measuring different schemas, scored by different rules.
Related course: DS103 — SQL for Data Analysis
Long Context Didn't Kill RAG. It Killed Lazy RAG.
On accuracy, long context often wins. It's the other three columns of the comparison that decide which architecture you need.
Related course: AI402 — Retrieval & RAG Systems
Prompt Engineering Won't Survive — the Skill Underneath It Will
The tip lists were a patch for weak models. The skill underneath them — specifying an ambiguous problem precisely enough for someone else to execute it — was never really about prompts.
Related course: DS106 — AI-Assisted Data Science
Your RAG Pipeline Doesn't Have an Accuracy Problem — It Has an Evaluation Problem
The demo answered ten questions you chose. Production will answer ten thousand you never saw — and nothing in your pipeline can tell you when it starts getting them wrong.
Related course: AI402 — Retrieval & RAG Systems
Clustering Isn't Discovery — It's a Hypothesis You Still Have to Test
Four customer segments appear as a fact from data. It is one modeling choice among several that happened to produce four groups.
Related course: ML203 — Unsupervised Learning & Clustering
You Don't Need a Math PhD for Data Science — You Need to Stop Skipping the Boring Step
Career-changers over-study algorithms and under-practice the unglamorous work that actually determines whether an analysis holds up.
Related course: DS101 — Intro to Data Science
The Most Dangerous Phrase in Analytics Is "The Data Speaks for Itself"
Every chart embeds choices. Naming them is what makes an analysis defensible.
Related course: DS104 — Statistics for Data Science
Your AI's EDA Looks Right. Here's How to Tell When It Isn't.
Six failure patterns that pass every test except a human's.
Related course: DS105 — Exploratory Data Analysis
The 3 Kinds of Data Leakage That Make Your Model Lie to You
Why AI code-review tools will catch one of them and miss the other two.
Related course: DS203 — Feature Engineering & Pipelines