Architecture
Pre-computed JSON
no backend, no runtime training
Case Study
Live Educational Platform
Lab XGBoost
Interactive ML Education for ICU Clinicians
Hands-on platform where clinicians without ML background build XGBoost models step-by-step. Pick clinical features, watch ROC curves update real-time, read actual Python code beside it. Phase 1 is Feature Explorer with deliberately tempting bait features (data leakage, identifiers, clinically irrelevant). Phase 2 is Hyperparameter Sandbox to feel underfitting, overfitting, and the sweet spot. All models pre-computed in Python and exported to static JSON — browser only does lookups, no runtime training, no backend.
Visualization
D3.js Live ROC
instant feedback
Phases
Feature → Hyper
progressive ML pedagogy
Key Achievements
- Designed two-phase pedagogy that teaches feature selection and hyperparameter tuning through play.
- Built deliberate "bait features" that reveal data leakage and identifier traps when picked.
- Pre-computed thousands of XGBoost variants in Python, exported as static JSON for browser lookup.
- Live ROC curve animation with D3 for immediate feedback on every parameter change.
- Inline syntax-highlighted Python code with PrismJS so learners see the actual model in action.
- Sentry-instrumented for production observability without slowing the learning experience.
Tech Stack
12 Astro 6 Svelte 5 TypeScript D3.js animejs PrismJS Sentry XGBoost scikit-learn Python 3 Tailwind CSS MDX
- Client
- dr. Eka Satrio Putra, Sp.An-TI
- Domain
- xgboost.lab.ekasatrio.id
Gallery
Free 15-min consult
Want something like this?
Tell us your scope, stack, and timeline. We'll send a clear next step.