Simplifying English Learning Through AI
An AI-powered platform that combines intelligent vocabulary tools, voice-based tutoring, and community — built for learners across Southeast Asia who need more than flashcards.
Phone mockups: Home screen, Dictionary, AI Voice Tutor
Client
Peilin — Founder, Melon Labs
Scope
Product Strategy
AI Engineering
Mobile Development
Community Infrastructure
My Role
Led product strategy, AI architecture, and technical execution across five surfaces — from dictionary to mobile app — while coaching a first-time founder and mentoring a junior developer.
Most English learning tools stop at memorization
Learners don't just need definitions — they need context, conversation, and a way to actually use the words they're learning. When they hit a word they don't understand inside a definition, there's nowhere to go. They're stuck.
The options for adult learners across Thailand, Korea, and Southeast Asia are either gamified apps that don't go deep enough or classroom instruction that can't scale. Nothing sits in the middle.
Image: The gap between flashcard apps and classrooms
Three layers, one connected product
AI Dictionary
LLM-powered definitions with confidence scoring, tappable words inside definitions, and multi-purpose chat — vocabulary, idioms, grammar, and translation in one interface.
Voice AI Tutor
Guided conversation sessions — voice-only input by design. Multiple modes: vocabulary practice, debate, Q&A, and idiom role-play. One backend, many experiences.
Community
Where learners validate AI answers with real tutors. Integrated with Heartbeat, Stripe + Xendit for SEA payments, live workshops, and membership tiers.
Every word inside a definition is tappable
If a learner doesn't understand a word in the definition itself, they can instantly check it without leaving the card. A tooltip shows the simple explanation plus options to save or look up the full definition.
Confidence indicators use traffic-light colors — green for sourced, yellow for LLM-generated using a source, red for pure LLM knowledge. User testing showed percentages caused overthinking. Simple colors gave learners exactly what they needed.
Dictionary card with tappable words and confidence indicators
Two-tier loading: simple explanation first, web-sourced content in background
30s
Initial load time per word
9-11s
After two-tier optimization
12,000+
Pre-cached word definitions
AI tutor mid-session — voice input, live transcription
Voice-only input. By design, not by limitation.
The goal is to push learners toward actually speaking rather than hiding behind text. We tested it ourselves — even native English speakers had to genuinely think through grammar when forced to speak their answers.
Powered by Deepgram for speech-to-text and text-to-speech. Responses stream in real-time with audio playing as sentences complete, eliminating dead-air.
Guided Practice
8-phase session: pronunciation, meaning, usage, collocations, and sentence construction for a single word.
Debate
The AI takes a position and the learner argues back, building fluency under pressure.
Q&A
Open-ended questions about English usage, grammar, or context.
Idiom Role-Play
Trade roles using expressions in real conversational scenarios, shadowing before improvising.
Five surfaces. One developer. One coherent experience.
Built with React Native and Expo. The app brings everything together: AI voice tutor, dictionary, daily vocabulary discovery, and an onboarding flow that captures the learner's goal in three questions.
Early home screens overwhelmed users. We stripped it back to a single "Discover New Word" card and a floating AI button. Reducing cognitive load was what made users actually start learning.
iOS and Android — home screen with daily vocabulary card
The business model wasn't obvious — and that was the point.
We explored B2C subscriptions, freemium tiers, early bird memberships, and a token-based retention model. Eventually a B2B path emerged: AI tooling for existing English learning communities across SEA with 500K+ members running on WhatsApp with zero AI capability.
SEA Payment Reality
Less than 5% of people in Indonesia use credit cards. We integrated Xendit for e-wallets alongside Stripe for Western markets — a constraint most teams wouldn't catch until launch day.
Scope Discipline
New ideas landed every week. We called a hard MVP freeze: locked the screens, stopped design changes, set a structured timeline. That discipline got the app shipped.
Founder Growth
She went from "I have zero networking with IT people" to running a live product ecosystem across five surfaces. That transformation was the most valuable outcome.
What made this engagement difficult
Five surfaces, one developer
Web dictionary, community integrations, marketing site, mobile app on two platforms — all running in parallel on a startup budget. We sequenced so each surface built on the last.
Third-party platforms fought us
Heartbeat had no payment endpoints. Apple verification took weeks. Expo's free tier took 30 minutes per build. We designed around every limitation.
Latency was the silent killer
30 seconds per word lookup. Noticeable tutor delay. We optimized at every layer — two-tier loading, streaming, caching, batch pre-generation. Each shaved seconds. Collectively, they made the difference.
Three time zones, five months
New Zealand, the US, and India. Keeping momentum through platform blockers, app store delays, and the natural ups and downs of a first-time founder's journey.
From idea to multi-surface product ecosystem
Live AI dictionary with confidence scoring, tappable lookups, and 12,000+ pre-cached definitions
AI voice tutor with 4 session modes, streaming responses, and live transcription
Community platform with Stripe + Xendit payments, membership tiers, and live workshops
Marketing website with newsletter automation, blog, events, and separate app waitlist
Mobile app in beta on iOS and Android with optimized onboarding
Full codebase, documentation, and accounts transferred to the founder