Airo: On-Device AI + RAG + Function Calling Product Strategy

Status: Phase 0 Complete ✅

Phase 1 Starting 🚀

Executive Summary

Airo is an AI-powered super app that processes PDFs, images, and audio on-device using Gemma 1B (quantized int4) with RAG and function calling. Zero network required for core workflows. Target platforms: Android (Pixel 9), iOS (iPhone 13 Pro Max), Chrome (desktop).

Key Differentiator: All data processing happens locally. No PHI leaves the device unless explicitly opted in.

Success Metrics (Implementable)

Metric	Target	Rationale
Offline Success Rate	90% of workflows	PDF→extraction→action with no network
Latency	<3s for 1-page PDF	Pixel 9 + Gemma 1B int4
App Footprint	<1.2GB	Model + app minimal install
Extraction F1	≥0.9	Bills, diet plans held-out test set
Battery	<5% per workflow	Single full run on Pixel 9

High-Level Architecture

┌─────────────────────────────────────────────────────────────┐
│                    Flutter UI (Cross-Platform)              │
│  (Login, File Upload, Results, Notifications, Payments)    │
└────────────────────┬────────────────────────────────────────┘
                     │ Platform Channels
        ┌────────────┼────────────┐
        │            │            │
    ┌───▼──┐    ┌───▼──┐    ┌───▼──┐
    │Android│   │ iOS  │   │Chrome │
    │Native │   │Swift │   │WASM   │
    └───┬──┘    └───┬──┘    └───┬──┘
        │           │           │
    ┌───▼───────────▼───────────▼──────────────────┐
    │  AI Edge SDK / LiteRT Runtime                │
    │  ├─ Gemma 1B int4 (Inference)               │
    │  ├─ Embedding Model (RAG)                   │
    │  ├─ Function Calling SDK                    │
    │  └─ OCR / PDF Parser                        │
    └───┬──────────────────────────────────────────┘
        │
    ┌───▼──────────────────────────────────────────┐
    │  Local Storage (Encrypted)                   │
    │  ├─ SQLCipher (Extracted Data)              │
    │  ├─ Vector Index (HNSW / AI Edge RAG)       │
    │  └─ Notification Schedule / Payment Requests│
    └────────────────────────────────────────────┘

Core Use Cases & Functions

1. fill_form(patient) - Healthcare/Diet Forms

Extract: name, DOB, weight, height, diet plan (day→meals→times→notes)
Output: Structured form data ready for submission
Validation: Required fields enforced

2. schedule_notifications(plan_id) - Recurring Reminders

Input: plan_id, start_date, recurrence (daily/weekly/custom), times
Action: Create local notification schedule (Android WorkManager / iOS UserNotifications)
Persistence: Store in local DB

Input: bill_id, participants (user_id, share_pct), items (name, price, assigned_to)
Output: Per-person shares, payment requests
Action: Store in local DB, optionally sync when online

Implementation Roadmap (10 Phases)

Phase 0: Foundation ✅ COMPLETE

Project structure, authentication, cross-platform setup
App name: “Airo” with AI+Air icon
Platforms: Android (Pixel 9), Web (Chrome), iOS config ready

Phase 1: PoC - On-Device AI (2-3 sprints)

Load Gemma 1B int4 on Android
Basic inference + fill_form function
Flutter platform channel bridge

Phase 2: Document Ingestion (2 sprints)

File picker, PDF/image/audio preprocessing
OCR (ML Kit + Tesseract fallback)
Rule-based field extraction

Phase 3: RAG & Vector Store (2 sprints)

Chunking, on-device embeddings, HNSW index
Retrieval pipeline, end-to-end RAG

Phase 4: Function Calling (1-2 sprints)

All three functions: fill_form, schedule_notifications, split_bill
Native handlers, local actions

Phase 5: iOS & Web Parity (2-3 sprints)

iOS: LiteRT/CoreML conversion, Swift native layer
Web: LiteRT Web/WASM, JavaScript RAG

Phase 6: Privacy & Security (1 sprint)

SQLCipher encryption, consent flows, data purge

Phase 7: Performance (1 sprint)

Latency, memory, battery optimization

Phase 8: Testing & Eval (2 sprints)

500 bills, 200 diet PDFs, 100 receipts
F1 ≥0.9 extraction accuracy

Phase 9: Beta (2 sprints)

Firebase App Distribution, user feedback, iteration

Phase 10: Production (Ongoing)

Play Store, App Store, web release
Monitoring, model updates, A/B testing

Technology Stack

Component	Technology	Notes
Model	Gemma 3 1B int4	LiteRT quantized, <500MB
Inference	AI Edge SDK / LiteRT	Android native, iOS CoreML, Web WASM
RAG	AI Edge RAG SDK	On-device retrieval + chunking
Function Calling	AI Edge Function Calling	JSON schema-based
OCR	ML Kit + Tesseract	Local, no cloud
Storage	SQLCipher + HNSW	Encrypted vectors + metadata
Notifications	WorkManager (Android) / UserNotifications (iOS)	Local scheduling
UI	Flutter	Single codebase, platform channels

Risk Mitigations

Risk	Mitigation
OOM on older devices	Smaller model, streaming tokens, cloud fallback
Function hallucination	Strict JSON schema, deterministic validators, human confirmation for payments
Multimodal sync issues	Prefer deterministic OCR for numeric/date data
Model accuracy drift	A/B testing, continuous evaluation, model updates

Next Steps (Week 1)

Download & quantize Gemma 1B → int4 LiteRT model
Set up Android native module → Load model, basic inference
Create Flutter plugin → Platform channel bridge
Register fill_form function → Test with sample form
Build Flutter UI → Text input, model response display

References

Product Manager: Airo Team
CTO: Architecture & Technology Lead
Last Updated: 2025-10-30