Back to Product

AI selection architecture document

AI Selection Architecture Document

Project Name: Duory - AI Language Learning Companion
Version: 1.0
Date: October 26, 2023

1. Introduction

Duory is a mobile-first language learning tool integrated with Duolingo, offering translation, romanization, and kana support. This document outlines the AI architecture for real-time language processing, focusing on scalability (target: 50K+ concurrent users), low-latency responses (<500ms), and offline capability.

2. AI Feature Requirements

Feature AI Requirement Criticality
Translation Multilingual NLP (50+ languages) High
Romanization Phonetic transcription (e.g., Japanese → Romaji) Medium
Kana Support Kanji-to-Kana conversion (Japanese) High
Offline Mode On-device lightweight ML models Medium
Adaptive Learning User progress analytics for personalized tips Low

3. Technology Selection & Justification

3.1 Core AI Frameworks

  • Translation & NLP:
    • Google Cloud Translation API v3 (Dynamic batch processing, 110+ languages, $20/million chars).
    • Backup for Offline: ONNX Runtime with distilled mBART-50 model (45MB size, 85% accuracy offline).
  • Romanization/Kana Conversion:
    • PyKakasi v2.0 (Python) for Japanese romanization (Apache 2.0 license).
    • Kuromoji (v0.9.0) for Kanji-to-Kana (Java/Kotlin).
  • Analytics:
    • TensorFlow Lite (v2.10) for on-device learning pattern detection.

3.2 Infrastructure

  • Cloud: Google Cloud Run (auto-scaling, cold-start <1s).
  • Mobile: React Native (v0.70) with ML model caching via React Native FS.
  • API Gateway: Apigee X (rate limiting, OAuth 2.0).

4. Architecture Overview

┌─────────────┐       ┌───────────────┐       ┌──────────────────┐  
│ Mobile App  │──────▶│ API Gateway   │──────▶│ Cloud Translation│  
│ (React      │       │ (Apigee X)    │       │ & PyKakasi       │  
│ Native)     │◀──────┤               │◀──────┤ (GCP Cloud Run)  │  
└─────────────┘       └───────────────┘       └──────────────────┘  
       ▲  │                     │                     │  
       │  └────▶ On-Device ─────┘                     │  
       │        Models (mBART-50)                     ▼  
       └────────── User Data Sync ───────────▶ Firestore DB (Encrypted)

5. Implementation Steps

Phase 1: Core AI Pipeline (4 Weeks)

  1. Integrate Cloud Translation API:
    • Implement batch translation with fallback to mBART-50 when offline.
    • Optimize payloads using Protocol Buffers.
  2. Embed PyKakasi/Kuromoji:
    • Compile to WebAssembly for React Native (via Emscripten).
    • Benchmark: <100ms per 100-character conversion.
  3. Model Caching:
    • Pre-load mBART-50 during app install using React Native FS.

Phase 2: Offline & Security (3 Weeks)

  1. Data Encryption:
    • Encrypt user logs using AES-256 (Android Keystore/iOS Keychain).
  2. Token-Based Access:
    • Validate Duolingo API tokens via OAuth 2.0 PKCE.
  3. Edge Caching:
    • Cache frequent translations via Cloud CDN (1-hour TTL).

Phase 3: Scalability (Ongoing)

  • Auto-scale Cloud Run instances (max 100 pods).
  • Monitor latency via Cloud Trace; SLO: 95% requests <500ms.

6. Security & Compliance

  • Data Privacy: GDPR/CCPA compliance; anonymize user data in analytics.
  • Threat Mitigation:
    • DDoS: Cloud Armor (rate limiting @ 1K reqs/user/minute).
    • Model Tampering: JWT-signed model updates.

7. Performance Metrics

Component Target
Translation Latency ≤400ms (online), ≤800ms (offline)
Romanization ≤100ms
API Uptime 99.95% (SLA-backed)
Model Size <50MB (on-device)

8. Future Extensibility

  • Voice Recognition: Add Whisper.cpp (offline ASR) in v2.0.
  • Personalization: Federated learning for adaptive content (TensorFlow Federated).
  • Multi-Platform: Export PyKakasi to WebAssembly for web support.

Approvals:

  • Lead Architect: [Signature]
  • AI/ML Engineer: [Signature]
    Document Length: 3,200 characters