CUNNY.
The Vision
CUNNY is a premium gamified AI learning platform for K-12 students, built on academic AI literacy research (AAAI-25/IEEE). Designed like Duolingo for AI, it combines 19 interactive sandbox widgets — where students manipulate neural network weights, simulate diffusion models, detect dataset bias, and evaluate AI prompts — with on-device machine learning (LiteRT/TFLite) for real-time fruit image classification with an Explainability Engine. The app features an offline-first architecture with Room DB caching and WorkManager cloud sync, a Glassmorphism 2.0 UI system with animated fluid backgrounds, retro PCM synthesizer audio feedback, and a full serverless backend on Cloudflare Workers (Hono + Neon PostgreSQL). Originally conceived as a capstone project at Bangkit Academy 2024, the application was independently expanded and rebuilt into a full-stack production-ready platform through 2026.
Development Roadmap
Phase 01: Research & Curriculum Design
Surveying Q1 AI literacy journals (AAAI-25, IEEE, Chiu & Sanusi) and structuring a bilingual curriculum across 4 categories (AI, Coding, Data, CNN) with 7 courses and 20 lessons aligned to academic pedagogy standards.
Phase 02: Serverless Backend API
Architecting and deploying the Content API on Cloudflare Workers using Hono.js with Neon PostgreSQL via Hyperdrive connection pooling. Implementing 9 route modules covering authentication (JWKS/JWT), course content delivery, student progress tracking, classroom management, and parent consent flows.
Phase 03: Jetpack Compose UI & Glassmorphism 2.0
Building 13 Compose screens with a custom 3-depth glassmorphism card hierarchy (standard/deep/light glass), animated fluid background with organic plum-colored orbs, floating pill navigation bar with spring physics, and Phosphor icon system — all adhering to the Warm Plum color palette.
Phase 04: 19 Interactive AI Sandbox Widgets
Developing 19 self-contained interactive simulations embedded within lessons: Neuron Sandbox (LTU weight manipulation), Diffusion Sandbox (noise-to-image), Bias Game (dataset bias detection), Prompt Evaluator (hallucination detection), Reward Trainer (reinforcement learning), Next-Word Predictor (language model probability), and 13 others.
Phase 05: On-Device ML & Explainability Engine
Integrating a bundled TFLite fruit classification model (~4MB, 20 classes) via LiteRT 1.4.2 for 100% offline inference. Building the image processing pipeline (bitmap resize, normalize, softmax) and Explainability Engine that surfaces top-K predictions with natural-language reasoning for why the AI chose its answer.
Phase 06: Offline-First Architecture & Gamification
Implementing Room database caching for lessons, bookmarks, and gamification state. Building XP (+10/lesson), daily streak tracking, energy system, and badge progression (Explorer Bronze → Completion Gold). Adding WorkManager-based sync queue with 3-retry exponential backoff for cloud synchronization and Lottie mascot "Coji" celebration animations with confetti.
Phase 07: Release Engineering & Compliance
Enabling R8/ProGuard code shrinking, verifying COPPA compliance for ages 13+, configuring Firebase Crashlytics monitoring, writing privacy policy and terms of service, generating signed release AAB, and preparing Play Store listing assets for public alpha submission.