Interactive Machine Learning Visualisation Platform
Back to projectsMachine learning & numerical methodsRegression, classification, Naive Bayes and decision trees implemented from scratch, so every lab can show its intermediate values.TypeScript Frontend & interactive visualisationA React app of draggable SVG charts and lazy-loaded 3-D scenes, with lessons written in MDX and progress saved in the browser.React 19 · TypeScript · Vite 8 · MDX 3 · KaTeX · three.js · react-three-fiber · SVG
Interactive Machine Learning Visualisation Platform
In DevelopmentReactTypeScriptthree.jsMDXKaTeXSVG
An interactive study site for introductory machine learning, with 65 lessons and 100 labs where every algorithm is written from scratch and re-runs live as you drag the data.
- Role
- Solo developer
- Period
- Sep 2026 – Present
Machine learning & numerical methods
- Regression solvers written from scratch: least squares through QR decomposition, ridge regression, and LASSO by coordinate descent with warm starts, on plain arrays so the labs can display each step. A seeded random number generator keeps the simulations reproducible.
- Classification and evaluation from first principles: logistic-regression loss and gradients, Minkowski distances, k-NN voting with explicit tie rules and distance weighting, leave-one-out error, confusion matrices, and ROC curves with AUC.
- Decision trees you can grow and prune: entropy, information gain and gain ratio, greedy classification trees with reduced-error pruning, and regression trees, so a learner can build a tree split by split and watch validation error change.
- Worked examples reproduced exactly: the labs recompute the standard worked examples, e.g. AUC = 7/9 and information gains of 0.266 vs 0.121. The Naive Bayes labs count with exact fractions instead of floats, so posteriors come out as 1/189 vs 18/875.
Frontend & interactive visualisation
- 100 labs on one reusable chart kit: a small SVG chart kit (axes, class markers, decision regions, contours, arrows) that every lab builds on. Dragging a point or moving a slider re-runs the algorithm and redraws the fit, boundary or loss surface.
- 3-D views that cost nothing until used: least squares as a projection, the plane fit and the loss bowl are three.js scenes on react-three-fiber. The ~250 kB library only downloads when a scene comes near the screen, and scenes render on demand, so an idle view does no work.
- Lessons written in Markdown, wired up at build time: lessons are MDX. Custom remark plugins turn $…$ maths into KaTeX, ## headings into sections, and count each page's words, labs and checks for its study-time estimate. A new week is a folder copied from a template, and routes, sidebar, pager and progress pick it up automatically.
- Progress that survives restructuring: lessons marked understood, quiz answers, practice ratings and notes are kept in localStorage, and nothing leaves the browser. Answers are keyed by lesson slug, and a one-time migration moved older saves over when lessons were renumbered.
- Usable on a phone, offline and on paper: the layout goes from three columns to a drawer and a bottom bar on phones, and labs restack with container queries. There are several colour themes, each week has a printable reference sheet, and a single-file build opens straight from disk with no server.






