DEV Community

Harshit Sharma
Harshit Sharma

Posted on

Touch Grass

#devchallenge #hf26challenge

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Language Tutor — A local-first AI language tutor that runs entirely in your terminal. Practice 15+ languages with complete privacy — all conversations, vocabulary, and progress data stays on your machine in SQLite. Only API calls leave your device for AI inference.

Built for: My friend who wants to learn Spanish privately without subscriptions, data tracking, or cloud sync. They needed something that works offline for review, respects privacy, and costs $0.

Problems solved:

  • ✅ No monthly fees — uses NVIDIA's generous free API tier (14,400 req/day)
  • ✅ Works offline — review vocab, view progress, export data without internet
  • ✅ Zero data leaves your machine except API prompts — SQLite stored locally
  • ✅ Customizable — 4 tutor personalities, 6 CEFR levels, 8 AI models
  • ✅ Portable — single binary, JSON export/import for backup or sharing

Demo

Terminal Demo (ASCII Recording) 

┌─────────────────────────────────────────────────────────────────┐
│ 🗣️  Language Tutor                    v0.1.0                  │
│ Step 1/3: Setup                                                   │
├─────────────────────────────────────────────────────────────────┤
│ 1. Add API Key                                                    │
│ 2. Pick Language                                                  │
│ 3. Start Chatting                                                 │
│                                                                   │
│ 🗣️  Language Tutor                                                │
│ Practice any language with AI — 100% private, runs in terminal   │
│                                                                   │
│ Step 1 of 2: Get your free API key                                │
│ 1. Open https://build.nvidia.com in your browser                 │
│ 2. Sign in (Google/GitHub/Email) — it's free                     │
│ 3. Click "Get API Key" and copy it                               │
│ 4. Paste it below and press Enter                                 │
│                                                                   │
│ ❯ nvapi-xxxxxxxxxxxx█                                             │
│                                                                   │
│ ✓ Connected! Press Enter to continue.                             │
│ Your key is saved locally. Only used for AI responses.           │
└─────────────────────────────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode 
┌─────────────────────────────────────────────────────────────────┐
│ 🗣️  Language Tutor                    v0.1.0                  │
│ Step 2/3: Chat                                                    │
├─────────────────────────────────────────────────────────────────┤
│ 💬 Chat                                          Spanish │ encouraging │ A1 │
├─────────────────────────────────────────────────────────────────┤
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ You                                                         │ │
│ │ Hola, ¿cómo estás?                                          │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ 🤖 Tutor                                                    │ │
│ │ ¡Hola! Estoy muy bien, gracias por preguntar. ¿Y tú?        │ │
│ │ (Hello! I'm very well, thanks for asking. And you?)         │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ You                                                         │ │
│ │ Estoy bien. Quiero aprender español para viajar.            │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ 🤖 Tutor                                                    │ │
│ │ ¡Excelente! (Excellent!) Viajar es una gran razón.          │ │
│ │ ¿Qué países hispanohablantes quieres visitar?               │ │
│ └─────────────────────────────────────────────────────────────┘ │
│ ❯ Estoy bien. Quiero aprender español para viajar.█             │
├─────────────────────────────────────────────────────────────────┤
│ Ctrl+Shift+H for shortcuts  |  Ctrl+1-5 to navigate  |  Esc for chat │
└─────────────────────────────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

Key Features in Action

Feature Demo
Vocabulary Review (SM-2) Ctrl+2 → Grade words with 0/1/3/5 (Again/Hard/Good/Easy)
Scenarios Ctrl+3 → Restaurant, Travel, Interview, Doctor, Emergency + Custom
Progress Dashboard Ctrl+4 → Streaks, heatmap, stats, Ctrl+E export, Ctrl+I import
Settings Ctrl+5 → Model, Personality, Difficulty, API Key, Simple Mode
Help Ctrl+Shift+H → All keyboard shortcuts

Code

Repository: D:/project/hacktoberfest/language-tutor (ready to push to GitHub)

Project Structure 

language-tutor/
├── src/
│   ├── index.ts              # CLI entry (Commander.js + TTY check)
│   ├── app.tsx               # Main Ink app with screen routing
│   ├── screens/
│   │   ├── Welcome.tsx       # 3-step onboarding (API → Language → Chat)
│   │   ├── Chat.tsx          # Streaming chat with message bubbles
│   │   ├── VocabReview.tsx   # SM-2 spaced repetition UI
│   │   ├── Scenarios.tsx     # 8 roleplay scenarios + custom
│   │   ├── Progress.tsx      # Streaks, stats, heatmap, export/import
│   │   └── Settings.tsx      # 4 tabs: Chat, Learning, Account, Advanced
│   ├── components/
│   │   ├── Header.tsx        # Step indicator (like berwapp)
│   │   └── MessageBubble.tsx # User/assistant message rendering
│   ├── services/
│   │   ├── llm.ts            # NVIDIA API client (streaming + non-stream)
│   │   ├── conversation.ts   # Chat persistence & history
│   │   ├── vocabulary.ts     # Vocab CRUD + SM-2 SRS algorithm
│   │   ├── progress.ts       # Analytics + JSON export/import
│   │   └── prompts.ts        # System prompt builder (personalities + CEFR)
│   └── utils/
│       ├── db.ts             # SQLite + schema (7 tables, WAL mode)
│       └── paths.ts          # Cross-platform data directories
├── templates/prompts/        # 15 language-specific system prompts
├── dist/                     # Built output (single ESM file)
├── package.json
├── tsconfig.json
├── tsup.config.ts
├── README.md
├── ARCHITECTURE.md
├── PLAN.md
└── TRACKER.md
Enter fullscreen mode Exit fullscreen mode

Installation & Run 

# Clone
git clone <your-repo-url>
cd language-tutor

# Install & build
npm install
npm run build

# Run
node dist/index.js
# or after npm publish:
npx language-tutor
Enter fullscreen mode Exit fullscreen mode

Tech Stack

Layer Technology
TUI Framework Ink 5.x (React for CLI)
CLI Parser Commander.js 12.x
Database better-sqlite3 11.x (WAL mode)
AI API NVIDIA (Nemotron, Llama 3.1, Mistral, Gemma, Phi)
Streaming Native fetch + SSE
Build tsup 8.x (ESM, Node 18+)
Language TypeScript 5.x (strict)

How I Built It

Open-Source AI Used

NVIDIA's Free API Tier — 8 open-weight models hosted free:

  • nvidia/nemotron-3.5-lightning-30b-a3b (default, fast)
  • meta/llama-3.1-8b-instruct / 70b / 405b-instruct
  • mistralai/mistral-large-2-instruct / mistral-nemo-12b-instruct
  • google/gemma-2-9b-it
  • microsoft/phi-3.5-mini-instruct

No local inference needed — runs on any machine with Node.js 18+ and internet for chat.

Architecture Highlights

  1. Local-First by Default — SQLite with WAL mode stores everything: conversations, messages, vocabulary, SRS state, daily progress, settings. Zero cloud sync.

  2. Streaming TUI — Ink (React for terminals) renders tokens in real-time via async generators consuming NVIDIA's SSE stream.

  3. SM-2 Spaced Repetition — Implemented from scratch in vocabulary.ts:

   // Ease factor update: EF' = EF + (0.1 - (5-q)*(0.08 + (5-q)*0.02))
   // Interval: 1, 6, then EF * previous_interval
Enter fullscreen mode Exit fullscreen mode

  1. Language-Specific Prompts — 15 .md templates in templates/prompts/ with:

    • Strict language enforcement ("IMPORTANT: Always respond in Spanish")
    • Common learner mistakes per language
    • Cultural context (formal/informal address, regional variations)

  2. Privacy-First Data Export — Full JSON round-trip for backup, machine transfer, or sharing with a friend.

Why Does Open Innovation Matter?

Open-weight models + free hosting = Democratized AI access.

Without NVIDIA's free tier (and similar offerings from Groq, OpenRouter, Together AI):

  • This project would need a $2,000+ GPU for local inference
  • Or $20+/month for commercial APIs (OpenAI, Anthropic)
  • My friend couldn't afford either

What open innovation made possible:

  • ✅ Zero cost to run — 14,400 free requests/day covers heavy daily use
  • ✅ Model choice — Switch between 8 models for speed vs quality
  • ✅ No vendor lock-in — Swap API endpoint in one file
  • ✅ Community-driven — Prompt templates are plain Markdown, anyone can contribute a language
  • ✅ Transparency — You know exactly what data leaves your machine (just the chat context)

Closed APIs would have meant: Usage limits, billing surprises, data retention policies, and no offline capability.

My Agent Session

This project was built with Claude Code (Anthropic's CLI agent). The development session included:

  • Project planning & architecture (PLAN.md, ARCHITECTURE.md)
  • Daily progress tracking (TRACKER.md)
  • Full implementation across 4 days
  • TypeScript strict-mode fixes
  • 15 language prompt templates
  • Build verification

Agent session available via DevRelay if needed for judging.

Submission Checklist

  • ✅ Working demo — node dist/index.js launches TUI immediately
  • ✅ Privacy-first — All data in local SQLite (%APPDATA%/language-tutor/)
  • ✅ Free to run — NVIDIA free tier, no subscriptions
  • ✅ Offline-capable — Vocab review, progress, export work without internet
  • ✅ 15 languages — ES, FR, DE, IT, PT, JA, KO, ZH, RU, AR, HI, TR, NL, PL, SV
  • ✅ 8 AI models — Llama, Mistral, Gemma, Phi, Nemotron via NVIDIA
  • ✅ SM-2 SRS — Proven spaced repetition algorithm
  • ✅ Export/Import — Full JSON portability
  • ✅ Keyboard-driven — Ctrl+1-5, Esc, Ctrl+Shift+H, no mouse needed
  • ✅ Simple Mode — Beginner-friendly (Chat + Vocab only)
  • ✅ Builds clean — npm run build → single ESM file, no TypeScript errors
  • ✅ MIT License — In package.json
  • ✅ README + Docs — Complete usage guide

Built with ❤️ for language learners everywhere — especially the friend who inspired this.

Hacktoberfest 2026 • Build for a Friend • Local-First AI

Top comments (0)