---

Projects

Magic Cube Solver

Six search algorithms racing a 5×5×5 diagonal magic cube — replayed move by move.

A 5×5×5 diagonal magic cube has 125 cells that must be a permutation of 1..125, with all 109 rows, columns, pillars and diagonals summing to 315. Six local-search and metaheuristic algorithms attack it server-side in Go, and the browser replays the resulting swap log on a scrubber against a live objective-value plot.

The Genetic Algorithm is the exception. Being a population method it has no single sequence of swaps, so it returns no trajectory — the interface charts it rather than animating it.

Algorithms

Steepest Ascent Hill Climbing · Hill Climbing with Sideways Move · Random Restart Hill Climbing · Stochastic Hill Climbing · Simulated Annealing · Genetic Algorithm

Built with

Go   Fiber   SvelteKit   Three.js   Chart.js   Tailwind   Vite   Docker

Links — Live demo · Front end · Solver API · Report (PDF, Indonesian)

---

Phishing URL Classifier

Two algorithms, each written twice — and the baseline that keeps them honest.

The UCI PhiUSIIL corpus runs 140,404 URLs deep and only 7.5% of it is hostile — which is exactly why accuracy flatters here and recall does the real work. KNN and Gaussian Naive Bayes are each built twice across its 49 features, once from scratch and once from scikit-learn, so every number has a reference to answer to. Around them the pipeline runs EDA, missing-value handling, outlier treatment, feature engineering, scaling and SMOTE.

Dataset: UCI PhiUSIIL Phishing URL Dataset (repository ID 967). Class 0 is phishing and is the positive class throughout.

Algorithms

K-Nearest Neighbors (from scratch) · K-Nearest Neighbors (scikit-learn) · Gaussian Naive Bayes (from scratch) · Gaussian Naive Bayes (scikit-learn) · EDA · Feature Engineering · SMOTE

Results — measured on a 28,081-row validation split.

Read every accuracy against that 0.9248 baseline. The corpus is 92.48% legitimate, so answering "legitimate" to everything scores 0.9248 while catching no phishing at all — which is why phishing recall, not accuracy, is the number that actually moves. The held-out file shipped with the dataset has no labels, so there is no test score and none is claimed.

This is a coursework reimplementation, not a security product. It is trained on a static 2023–24 dataset, has no threat intelligence, no blocklist, and no knowledge of any campaign newer than its training data. Do not use it to decide whether a link is safe.

Built with

Python   scikit-learn   NumPy   pandas   SciPy   Streamlit   Jupyter   Docker

Links — Live demo · Repository · Report (PDF, Indonesian)

---

Team

---