Spam Classifier Shipped — 5,572 Messages, Zero Mock — Hariom Lohar

Ship, don't demo

spam_classifier is now a live bench, not a screenshot. Type a message → see tokens split, scored red/green, weak signals filtered, then a SPAM / HAM stamp with P(spam|message).

Data: 5,572 rows, data.csv fetched in-browser
Engine: Naive Bayes + Laplace k=1, trained via Pyodide (Python in WASM)
UX: bench paper sheets, running verdict, gauge, weak-signal filter

Try it: /projects/spam_classifier.html

What I learned

Shipping in the browser surface uncovered what a notebook hides:

Tokenisation is a design choice, not a step
Thresholds are product, not math — 0.5 is a starting line, not an answer
If you can't show P(spam|message) with real numbers, you didn't teach it

Lab note

Every shipped artifact here gets a bench. This one proves the loop: derive → build → let the user touch it.

— Hariom · Lab Notebook No.01