Case study · Voice AI · MelonLabs

WordBuddy AI

Shipped to Google Play with 80% lower conversational-AI costs.

WordBuddy AI is an AI-powered English learning platform for Southeast Asian learners: an AI dictionary, voice AI tutors and gamified spaced repetition across 10,000+ CEFR, TOEIC, Business and GRE words.

Client: MelonLabs Ltd
Role: Full-Stack AI Engineer, leading backend and voice AI
Timeline: Mar 2026 – present

Book a call Visit the live product

The challenge

WordBuddy started as a web app. To reach learners where they actually study, it had to become a real mobile product: one that works offline, keeps voice tutoring affordable as usage grows, and never loses a learner's progress.

What I built

Evolved the web app into a cross-platform mobile app on Capacitor and Phaser.js, designed for offline-first use and game-driven learning.
Architected MelonTutor, the real-time conversational English tutor, on LiveKit, Deepgram, Gemini and ElevenLabs.
Re-architected storage: a 5 MB localStorage limit that silently lost data past about 1,000 saved words was replaced with encrypted on-device SQLite (SQLCipher), an IndexedDB fallback on the web and a row-level delta-sync engine for offline and multi-device consistency.
Added intelligent response caching and moved search and LLM providers to Exa AI and Gemini, with no drop in answer quality.
Migrated hosting from Render and Vercel to a single Hostinger VPS with one deployment pipeline.
Owned the release end to end: database schema, automated subscription workflows, CI/CD and publishing on Google Play.