Machine Learning · Internship Month 01

Two classifiers.
One reads text, one reads pixels.

A spam filter trained on real email data, and a digit reader trained on 42,000 handwritten samples from Kaggle's MNIST set — built, evaluated, and tested end to end in Python.

live_check.py

97.0%

Spam detection accuracy

96.6%

Digit recognition accuracy

47.7K

Labelled samples trained on

CASE 01

Email Spam Classification

Every message boils down to one question: does the pattern of words look like something a person wrote, or something designed to get clicked? The model answers by reading term frequency, not by reading intent.

Pipeline

01Clean the raw text — lowercase, strip URLs, digits, and punctuation.
02Vectorize with TF-IDF across unigrams and bigrams, capped at 3,000 features.
03Train two candidates — Multinomial Naive Bayes and Logistic Regression — on an 80/20 stratified split.
04Score both on precision, recall and F1, then keep the stronger model.

Spec sheet

datasetKaggle · Ashfak Yeafi

samples5,572 messages

split4,825 ham / 747 spam

featuresTF-IDF · 3,000 dims

modelMultinomial Naive Bayes

accuracy97.0%

spam precision1.00

Confusion matrix for the spam classifier — fig. 01 — confusion matrix, best model

Model comparison chart — fig. 02 — Naive Bayes vs Logistic Regression

Sample readouts

"Congratulations! You've won a free iPhone, click here to claim now!!!"SPAM

"Hey, are we still meeting for lunch tomorrow?"HAM

"URGENT: your account will be suspended, verify your password immediately"SPAM

"Can you send me the report before 5pm?"HAM

Interactive mode — the script drops into a live prompt after training: type any message and get an instant spam/ham verdict from the trained model.

Try it live — runs the real trained model, in your browser

This calls the actual Naive Bayes weights and TF-IDF vocabulary exported from training — not a simulation.

CASE 02

Handwritten Digit Recognition

Every digit arrives as 784 numbers — one per pixel, brightness 0 to 255, no picture file involved. The model never sees a "3"; it sees a specific arrangement of light and dark, and learns which arrangements repeat.

Pipeline

01Flatten each 28×28 image into a 784-value vector and standardize pixel intensities.
02Split 42,000 labelled images 80/20, stratified across all ten digits.
03Train two candidates — an RBF-kernel SVM and a 200-tree Random Forest.
04Score both on accuracy and a full per-digit report, then keep the stronger model.

Spec sheet

datasetKaggle · Digit Recognizer

samples42,000 images

resolution28 × 28 px

features784 scaled pixels

modelRandom Forest · 200 trees

accuracy96.6%

most confused3 / 5, 4 / 9

Confusion matrix for the digit classifier — fig. 03 — confusion matrix, all ten digits

Sample digit predictions — fig. 04 — ten test images, predicted vs. actual

Interactive mode — pick any test image by index, or ask for a random one, and the script saves a picture of that exact digit next to the model's live prediction.

Try it live — 200 real held-out test images

Each one is an actual Kaggle MNIST image the model never trained on, with its real prediction from the trained Random Forest.

200 samples drawn from an 8,400-image held-out test set.

Two classifiers.One reads text, one reads pixels.

Email Spam Classification

Pipeline

Spec sheet

Sample readouts

Try it live — runs the real trained model, in your browser

Handwritten Digit Recognition

Pipeline

Spec sheet

Try it live — 200 real held-out test images

Built with

Two classifiers.
One reads text, one reads pixels.