Airo Quick Reference Guide

🎯 Project Overview

Airo = On-device AI + RAG + Function Calling for PDF/image/audio processing

Platforms: Android (Pixel 9), iOS (iPhone 13 Pro Max), Chrome (Web)

Core Functions:

fill_form() - Extract & populate healthcare/diet forms
schedule_notifications() - Create recurring reminders
split_bill() - Extract bills & compute expense splits

📊 Success Metrics

Metric	Target
Offline Success	90%
Latency	<3s per PDF
Accuracy (F1)	≥0.9
Battery	<5% per workflow
Footprint	<1.2GB

🏗️ Architecture at a Glance

Flutter UI
    ↓ (Platform Channels)
Android/iOS/Web Native Layer
    ↓
AI Edge SDK (Gemma 1B int4)
    ↓
Local Storage (SQLCipher + HNSW)

📋 Phase Breakdown

Phase	Duration	What	Status
0	✅	Foundation	DONE
1	2-3w	PoC AI	STARTING
2	2w	Ingestion	QUEUED
3	2w	RAG	QUEUED
4	1-2w	Functions	QUEUED
5	2-3w	iOS/Web	QUEUED
6	1w	Privacy	QUEUED
7	1w	Performance	QUEUED
8	2w	Testing	QUEUED
9	2w	Beta	QUEUED
10	∞	Production	QUEUED

🚀 Phase 1 Checklist (Next 2-3 Weeks)

Week 1: Model Setup

Download Gemma 3 1B from Hugging Face
Quantize to int4 using LiteRT
Verify <500MB size
Test on Pixel 9 emulator

Week 2: Android Native

Add AI Edge SDK to build.gradle
Create GemmaInferenceEngine.kt
Create FunctionCallingManager.kt
Test basic inference

Week 3: Flutter Bridge

Create ai_edge_bridge plugin
Implement platform channels
Build chat UI
Test end-to-end on Pixel 9

🛠️ Tech Stack

Component	Tech	Size
Model	Gemma 1B int4	<500MB
Inference	LiteRT	-
RAG	AI Edge RAG	-
Functions	AI Edge FC	-
OCR	ML Kit	-
Storage	SQLCipher	-
UI	Flutter	-

📁 Project Structure

airo_super_app/
├── app/                          # Main Flutter app
│   ├── lib/src/
│   │   ├── features/
│   │   │   ├── auth/            # Authentication
│   │   │   ├── ai/              # AI chat interface
│   │   │   ├── documents/       # File upload
│   │   │   ├── results/         # Display results
│   │   │   ├── notifications/   # Schedule reminders
│   │   │   └── payments/        # Bill splitting
│   │   └── core/                # Shared utilities
│   ├── android/
│   │   └── app/src/main/kotlin/com/airo/superapp/
│   │       └── ai/              # Native AI layer
│   ├── ios/
│   │   └── Runner/              # iOS native layer
│   └── web/                      # Web assets
├── packages/
│   ├── airo/                     # Airo package
│   ├── airomoney/               # AiroMoney package
│   └── ai_edge_bridge/          # Platform channel plugin
└── docs/
    ├── PRODUCT_STRATEGY.md
    ├── TECHNICAL_ARCHITECTURE.md
    ├── ENGINEERING_CHECKLIST.md
    └── EXECUTIVE_SUMMARY.md

🔑 Key Files to Know

File	Purpose
`app/pubspec.yaml`	Flutter dependencies
`app/android/app/build.gradle.kts`	Android config + AI Edge SDK
`app/lib/main.dart`	App entry point
`app/lib/src/features/ai/screens/chat_screen.dart`	Chat UI
`packages/ai_edge_bridge/`	Platform channel plugin
`app/android/app/src/main/kotlin/com/airo/superapp/ai/`	Android native AI

🔄 Data Flow: Bill Processing

1. User uploads PDF
   ↓
2. PDF text extraction (text layer or OCR)
   ↓
3. Rule-based field extraction (amounts, dates, items)
   ↓
4. Chunking + embedding (RAG context)
   ↓
5. LLM inference with context
   ↓
6. Function calling: split_bill()
   ↓
7. Compute shares, create payment requests
   ↓
8. Store in encrypted DB
   ↓
9. Display results in Flutter UI

📱 Platform Channels

Android Method Channel: com.airo.ai_edge_bridge

Methods:

init(modelConfigJson) → Initialize model
query(prompt, optsJson) → Get model response
executeFunction(functionName, argsJson) → Execute function
indexDocument(filePath, metaJson) → Index for RAG

🎯 Function Schemas

fill_form

{
  "name": "fill_form",
  "parameters": {
    "name": "string",
    "dob": "date",
    "weight_kg": "number",
    "height_cm": "number",
    "diet_plan": [{"day": "int", "meals": ["string"], "times": ["string"]}]
  }
}

schedule_notifications

{
  "name": "schedule_notifications",
  "parameters": {
    "plan_id": "string",
    "start_date": "date",
    "recurrence": "daily|weekly|custom",
    "times": ["HH:MM"]
  }
}

split_bill

{
  "name": "split_bill",
  "parameters": {
    "bill_id": "string",
    "participants": [{"user_id": "string", "share_pct": "number"}],
    "items": [{"name": "string", "price": "number", "assigned_to": ["string"]}]
  }
}

🔐 Privacy & Security

Encryption: SQLCipher for local DB
Consent: Explicit opt-in for cloud fallback
Data Purge: User-initiated deletion
No Telemetry: By default (opt-in only)
Keystore: Device keystore for encryption keys

📊 Testing Strategy

Dataset:

500 real-world bills
200 diet plan PDFs
100 scanned receipts

Metrics:

Extraction F1 ≥ 0.9
Function call accuracy ≥ 0.95
Latency <3s per PDF
90% offline success rate

🚨 Common Issues & Fixes

Issue	Fix
Model too slow	Use GPU delegate, streaming tokens
OOM on device	Smaller model, lazy loading
Function hallucination	Strict schema, human confirmation
OCR accuracy low	Combine with rule-based extraction
Network required	All ops work offline

📚 Documentation

EXECUTIVE_SUMMARY.md - High-level overview
PRODUCT_STRATEGY.md - Product vision & roadmap
TECHNICAL_ARCHITECTURE.md - System design & components
ENGINEERING_CHECKLIST.md - Phase-by-phase tasks
QUICK_REFERENCE.md - This file

🔗 External Resources

👥 Team Roles

Role	Responsibility
Product Manager	Vision, roadmap, success metrics
CTO	Architecture, tech decisions, performance
Android Lead	Native AI layer, platform channels
iOS Lead	Model conversion, Swift implementation
Web Lead	LiteRT Web, WASM, JavaScript
Flutter Lead	UI, state management, integration
ML Engineer	Model selection, quantization, evaluation

Last Updated: 2025-10-30
Version: 1.0