Agentic Mobile Map

Transform 5-6 manual steps into 1-2 conversational turns

An intelligent errand routing system powered by dual AI agents that understands natural language and autonomously optimizes multi-stop journeys.

---

The Problem

Traditional errand routing apps require a tedious manual process:

1. Set your anchor point (home/work)
2. Define search radius
3. Search for first stop
4. Select from options
5. Search for second stop
6. Select from options
7. Manually optimize route
8. Finally navigate

This 5-6 step process is frustrating and time-consuming.

The Agent-Powered Solution

Simply say: "Take me home via Starbucks and Walmart"

Our dual-agent architecture does the rest:

• Natural Language Understanding: Parse complex requests in one utterance
• Autonomous Entity Resolution: Resolve "home" to your saved location, find the best Starbucks and Walmart on your route
• Smart Route Optimization: Automatically order stops to minimize detours
• Dynamic Detour Budget: Calculate acceptable detour distance based on route length
• Confidence-Based UI: Different flows for high/medium/low confidence results

Dual-Agent Architecture

• Gemini 2.5 Pro (Fast Agent): Handles 85% of requests with quick intent classification and entity extraction
• Gemini 3.0 Pro (Advanced Agent): Handles complex disambiguation and multi-turn reasoning for the remaining 15%
• Intelligent Routing: Automatically escalates to advanced agent when confidence drops below 0.60

---

Tech Stack

Frontend

• Framework: React Native with Expo
• State Management: Redux Toolkit
• Language: TypeScript
• Maps: Google Maps SDK
• Offline Storage: SQLite (mobile) / IndexedDB (web)

Backend

• Framework: NestJS
• Database: PostgreSQL (main data) + Redis (caching)
• Queue: Bull (async job processing)
• Language: TypeScript

AI & APIs

• AI Agents: Gemini 2.5 Pro (fast) + Gemini 3.0 Pro (advanced)
• Maps: Google Maps API
• Places: Google Places API

---

Key Features (Planned)

Conversational Interface

• Chat-like UI for natural language input
• Three-tier confidence system (HIGH/MEDIUM/LOW)
• Context-aware responses

Intelligent Routing

• Dynamic detour budget calculation (5-10% of route length)
• Stop status classification (NO_DETOUR, MINIMAL, ACCEPTABLE, NOT_RECOMMENDED)
• TSP-based route optimization

Smart Place Search

• Relevance-based ranking
• Automatic disambiguation
• Proactive route suggestions

Offline Support

• Cache top 20 popular stops near saved anchors
• Store last 10 routes
• Non-blocking background sync

Route Adjustment

• Visual route preview with map
• Drag-and-drop stop reordering
• Add/remove stops after optimization

---

Project Status

Current Phase: Initial Development

• ✅ Project structure and architecture design
• ✅ Frontend foundation with Expo and Redux
• â�³ Backend API development (NestJS)
• â�³ Gemini agent integration
• â�³ Google Maps/Places integration
• â�³ UI component development

---

Example Flow

` User: "Take me home but stop at Chick-fil-A and Walmart on the way"

App: "Perfect! Both are on your way with minimal detour.

 Route:
 • Current Location
 • Chick-fil-A (2.1 miles, +0.3 mi detour)
 • Walmart (7.8 miles, no extra distance)
 • Home (10.2 miles total)
 
 Estimated time: 18 minutes
 Ready to navigate?"

User: "Yes"

[Opens navigation with all waypoints set] `

---

Success Metrics

• 95% of routes fit within distance budget without user adjustment
• 90% correct destination disambiguation on first try
• <2 seconds response time for route suggestions
• 85% accuracy on "best" stop selection
• Full offline functionality with cached data

---

Architecture Highlights

Three-Tier Confidence System

1. HIGH (≥0.80): Execute immediately, show results
2. MEDIUM (0.60-0.79): Show confirmation dialog
3. LOW (<0.60): Escalate to Gemini 3.0 Pro, offer alternatives

Detour Budget Algorithm

• Short routes (≤2 miles): 10% max detour
• Medium routes (2-10 miles): 7% max detour
• Long routes (>10 miles): 5% max detour
• Absolute bounds: 400m minimum, 1600m maximum

Caching Strategy

• Routes: 1 hour (traffic changes frequently)
• Places: 7 days (details relatively stable)
• Anchors: 30 days (user locations rarely change)
• Disambiguation: 14 days

---

Development

Frontend Setup

�ash cd frontend npm install npm start

Backend Setup (Coming Soon)

�ash cd backend npm install npm run start:dev

---

License

MIT License - see LICENSE file for details

---

What Makes This Different?

Most map apps make you do the work. We use AI agents to:

• Understand context: "on the way" means we calculate detour budgets automatically
• Resolve ambiguity: Multiple Starbucks? We pick the one that makes sense for your route
• Optimize autonomously: No manual "add stop" → "optimize" → "check again" loops
• Learn preferences: Your frequent stops become suggestions

The goal: Natural conversation, not manual configuration.