Your ML Model Does Not Understand Dates

How a sine-cosine pair, a decomposed timestamp, and a couple of other cheap transforms can move accuracy as much as the model itself

Your Fancy Recommender Loses to “Most Popular” Half the Time

I built collaborative filtering and matrix factorization from scratch, then raced them against the laziest baseline I could write: it held…

RAG 101: Retrieval Does Not Fix Hallucination, It Just Moves the Blame

I changed one setting nobody tunes and answer accuracy fell from 100 percent to 33 percent, even though the model never moved

LoRA From Scratch: Two Small Matrices Instead of a Million Weights

I hid a change of known size inside a small network, then watched a pair of thin matrices recover all of it while the rest of the model…

Random Sampling 101: A Million Draws Saw Nothing, and Five Thousand Found the Answer

Below are the places random sampling actually misleads you

Your Regression Model Gives You a Number, but It Should Give You a Range

A prediction interval makes a promise about how often it contains the truth, but almost nobody checks it

Optimization 101: When Your Business Problem is Really a Set of Equations

Write product mix as three lists, hand it to a free solver, and you get the exact best plan, plus a number no forecast will ever give you…

Backpropagation 101: 40 Lines of NumPy That Learn XOR

A tiny network with no framework learns a problem no straight line can solve, and every gradient it computes matches a numerical check to…

Marketing Mix Models 101: When Your TV Ad Did Not Return 21x, but Christmas Did

I built a synthetic market where I set the real returns myself, then watched an ordinary regression hand the whole TV budget credit for the…

Apache AGE 101: The Graph Database Hiding Inside Your Postgres

A four-hop question needs four self-joins in SQL and one line of Cypher in Apache AGE, so I built a small test bench to measure what that…

Developer Salaries Rose 5% Between Surveys, but the Same Developer Got 1%

The Stack Overflow salary survey median jumped 5% in a year, but hold the developer fixed and most of that raise disappears

Class Weights 101: The One-Line Fix for Imbalanced Data That Beats SMOTE

One keyword argument that also rewrites your predicted probabilities

I Built a Transformer From Scratch in NumPy, and It Learned to Sort by Looking in the Right Place

No GPU, no libraries, about 68,000 parameters you can read line by line

You Measured Distance as the Crow Flies, but Your Van Drives on Roads

In central Copenhagen the road trip runs 1.37 times the straight line on average, 3.4 times across the harbour, and the straight-line…

Your End-to-End Tests Do Not Need Step Definitions

Two agents and a real browser run a Gherkin file with no step definitions underneath it, and code checks every number

Six Outlier Detectors, One Dataset, and Barely Any Agreement

I planted the anomalies myself, so I knew every right answer, but the three-sigma rule missed most of them while six detectors flagged…

Market Basket Analysis Tells You What Sells, Not Who Is Leaving

A synthetic store with a known answer key, and what it says about which rules to trust and which customers are worth a retention offer

Hashing 101: Count Half a Million Users in 1.5 Kilobytes

One idea from 1970 lets you answer questions an exact set cannot afford, once you accept an error rate you choose up front

Your Pandas Code Is the Slowest Way to Do Almost Everything

Ten everyday operations, four engines, a two-core box, and the largest speedup I found had nothing to do with switching engines