Built an AI English Learning App with Custom Fine-Tuned LLM

The Idea

Learning English through AI should feel like talking to a real tutor — not a robotic chatbot. I wanted to build an app that understands grammar mistakes, explains corrections naturally, and adapts to the user's level.

So I trained my own LLM model specifically for English language teaching.

Training My Own LLM

Dataset Preparation

I curated a training dataset of 10,000+ English conversation examples including:

• Common grammar mistakes and their corrections
• Natural conversation flows at different proficiency levels
• Vocabulary building exercises with context
• Idioms and phrasal verbs explained simply

Fine-Tuning Process

Using Ollama and Runpod GPU instances, I fine-tuned a base model:

1. Prepared data in instruction/response JSONL format
2. Ran QLoRA fine-tuning on Runpod A100 GPUs
3. Evaluated against English proficiency benchmarks
4. Converted to Ollama format for deployment

The fine-tuned model outperformed the base model by 45% on grammar correction tasks.

The App

Built with React Native for both iOS and Android, the app features:

AI Chatbot

• Real-time conversation with the fine-tuned English tutor
• Instant grammar analysis and correction
• Natural explanations of why something is wrong
• Adapts difficulty based on user level

Progress Tracking

• Conversations — track daily chat sessions
• Study Time — monitor learning minutes
• Progress — overall improvement percentage
• Achievements — gamified milestones (10 earned so far!)

Beautiful UI

• Gradient-based design with warm orange/pink tones
• Bottom navigation: Home, Practice, Profile, Settings
• Progress cards with icons for quick stats
• "Start New Session" button for instant learning

Tech Stack

• React Native — Cross-platform mobile app
• TypeScript — Type-safe development
• Ollama — Custom fine-tuned LLM serving
• Express.js — RESTful API backend
• PostgreSQL — User data, progress, session history, grammar logs
• Firebase — Authentication and push notifications
• Runpod — GPU infrastructure for model training

Backend Architecture

The backend is built with Express.js and PostgreSQL:

Express.js API
├── /auth         — JWT authentication (signup, login, refresh)
├── /chat         — AI conversation endpoints (stream responses)
├── /grammar      — Grammar analysis & correction logs
├── /progress     — User stats, achievements, streaks
├── /sessions     — Learning session management
└── /models       — Model selection & configuration

PostgreSQL handles all persistent data:

• users — profiles, settings, subscription tier
• sessions — chat history with timestamps
• grammar_logs — corrections tracked for analytics
• achievements — unlocked milestones per user
• progress — daily/weekly/monthly stats

Key Features

Results

• 1,000+ conversations in the first month of testing
• 45% better grammar correction than base model
• 4.8/5 user satisfaction rating from beta testers
• 19% average progress improvement in 2 weeks

What I Learned

• Fine-tuning a model for a specific domain (English teaching) produces dramatically better results than using a general-purpose LLM
• Express.js + PostgreSQL gives you full control over data — Firebase alone isn't enough for complex analytics
• React Native with a proper backend scales much better than a pure Firebase approach
• Gamification (achievements, progress bars) keeps users engaged with learning
• The quality of training data matters more than the quantity

This project combines my passion for AI/ML with practical mobile development — exactly what I do at Glixen Tech.