Dealpilot AI

Inspiration

Online shopping has become easier, but the process is still fragmented. Finding the right product often involves switching between multiple apps, researching specifications, comparing prices, checking availability, and navigating complicated checkout processes.

We asked ourselves a simple question:

What if shopping could be as easy as having a conversation?

That question inspired DealPilot Agent, an AI-powered Android shopping assistant designed to connect product discovery, comparison, decision-making, and payments through one intelligent conversational interface.

Our vision was to move beyond traditional shopping chatbots and explore an autonomous AI agent capable of interacting with real Android applications, researching products, understanding user preferences, and connecting the resulting shopping journey with PayPal.

What It Does

DealPilot Agent transforms a shopping request into an AI-assisted mobile shopping workflow.

Instead of manually browsing multiple websites and comparing products, users can simply ask:

"Find the best wireless headphones under ₹10,000, compare their specifications, and help me buy the best option."

DealPilot Agent supports the following workflow:

AI-powered product discovery: The agent uses supported Android browsing tools to research products, examine available information, and identify relevant options.

Intelligent comparison: Users can request comparisons of prices, specifications, features, and product alternatives.

Conversational product selection: The user selects a product naturally, and DealPilot Agent retrieves the chosen model from the conversation rather than relying on a fixed product catalogue.

Cart preparation: The agent prepares a reviewable test cart containing the selected product, product reference, and user-confirmed checkout amount.

PayPal Sandbox checkout: When the user explicitly requests payment, DealPilot Agent connects to the PayPal Developer Platform and opens the official Sandbox approval experience.

Payment verification: The integration supports verification of approved Sandbox payments and presentation of capture and receipt details.

The goal is to keep product research and checkout connected while ensuring that the user remains in control of payment authorization.

How We Built It

DealPilot Agent combines a native Android application, AI model integrations, Android automation, and the PayPal Developer Platform.

AI and Agent Architecture

We developed DealPilot Agent around a conversational AI workflow that supports reasoning, task delegation, and tool execution.

The application supports OpenAI Codex account authentication and configurable AI-provider connections using supported API endpoints and model credentials.

The agent interprets user requests, coordinates supported tools, and returns findings through the main chat interface.

Android Automation with Shizuku

A major technical component is the integration of Android device-control capabilities.

We used:

Android Accessibility Services for supported screen observation and interaction.

Shizuku for additional Android capabilities and virtual-display operations.

MCP-based tool integrations to connect AI-driven workflows with available tools.

Native Android services for agent execution, conversations, and background-task management.

This enables DealPilot Agent to go beyond generating text and interact with supported phone interfaces while researching products.

PayPal Developer Platform Integration

We integrated PayPal Sandbox and the PayPal Orders API v2 into the checkout workflow.

The payment architecture uses a private companion backend to authenticate with PayPal, create Sandbox orders, handle approval, and verify payment status.

An important design decision was separating shopping actions from payment authorization.

Searching, comparing products, and preparing a cart cannot independently trigger a payment. Only an explicit user instruction such as "Proceed to pay" enters the payment workflow, followed by review and approval of the exact test amount.

The Sandbox environment allows us to develop and test payment interactions without transferring real funds.

Technology Stack

Mobile application: Native Android, Java, JavaScript, Android SDK

AI: OpenAI Codex and configurable AI model APIs

Agent orchestration: Embedded runtime and tool-driven task execution

Device automation: Shizuku, Android Accessibility Services, MCP integrations

Backend: Node.js, TypeScript, private checkout service

Payments: PayPal Developer Platform, OAuth 2.0, Orders API v2, PayPal Sandbox

Development and testing: Android Debug Bridge (ADB), GitHub, GitHub Actions, AI-assisted development tools

Challenges We Faced

Connecting AI reasoning with real Android actions

One of our biggest challenges was making an AI assistant interact with Android applications rather than simply generate shopping recommendations.

Android permissions, accessibility behavior, virtual displays, and asynchronous operations required careful integration and testing.

We addressed these challenges by retaining native device-control capabilities and incorporating Shizuku-based operations where supported.

Maintaining accurate product selection across conversations

During testing, we discovered that cart preparation could select a default product even when the user had chosen something different.

For example, selecting Sony headphones could incorrectly populate the cart with a JBL product.

We fixed this by implementing conversation-aware product selection that prioritizes the user's latest explicit choice and extracts matching product details from the chat.

We also added safeguards for ambiguous selections and uncertain prices, preventing the application from automatically inventing checkout amounts.

Integrating PayPal with an autonomous agent

Financial transactions require stricter control than ordinary browsing.

We needed to ensure that the AI could assist with shopping without initiating payments simply because a product was mentioned or selected.

We separated product discovery, cart preparation, payment initiation, and payment verification into distinct stages, with explicit buyer confirmation before opening PayPal Sandbox.

Testing on a physical Android device

Testing the application on an actual Android phone introduced challenges involving USB debugging, Accessibility permissions, Shizuku authorization, app signing, and communication with the checkout backend.

We performed Android installation tests, checkout-intent tests, and physical-device checks to validate key interactions.

The Sony headphone selection, cart preparation, and PayPal Sandbox payment-review stages were verified on the connected device. Full Sandbox buyer approval and receipt verification remain areas for additional end-to-end validation.

Accomplishments We're Proud Of

Our biggest accomplishment is building an Android AI assistant that combines conversational intelligence, real device interaction, and PayPal Sandbox checkout integration.

We are particularly proud of:

Building the shopping experience around one primary conversational interface.

Connecting AI-driven research with native Android browsing tools.

Supporting Shizuku and Android Accessibility for advanced device interaction.

Making cart preparation depend on the user's selected product rather than hardcoded listings.

Integrating PayPal Sandbox into an explicitly authorized checkout flow.

Building, installing, and testing the application on a physical Android device.

Publishing a downloadable Android APK and open-source code through GitHub.

The project demonstrates how AI agents can connect information gathering and transaction workflows while maintaining clear user approval boundaries.

What We Learned

This project taught us that building an autonomous AI application involves much more than connecting a language model to an interface.

We learned how to design tool-driven agent workflows, coordinate Android services, work with Shizuku and Accessibility permissions, manage state across conversations, and integrate secure payment APIs.

Working with the PayPal Developer Platform helped us understand OAuth authentication, order creation, buyer approval, payment capture, and the importance of transaction verification.

We also learned how valuable real-device testing is. Issues that were not apparent in isolated code tests became obvious when we exercised the full mobile interface.

Most importantly, we learned that AI autonomy and user control should complement each other, especially when a workflow involves payments.

What's Next for DealPilot Agent

Our next goal is to evolve DealPilot Agent into a more capable, reliable, and accessible mobile commerce assistant.

Future improvements include:

Integrating verified retailer product feeds for more reliable pricing and availability.

Supporting additional shopping platforms and merchants.

Improving natural-language understanding of product preferences and follow-up selections.

Reducing the setup required for Android device automation.

Moving the private checkout service to a securely hosted environment.

Completing broader end-to-end PayPal Sandbox payment and receipt validation.

Expanding voice-driven shopping and accessibility features.

Supporting more sophisticated AI-powered shopping workflows.

Our long-term vision is a mobile assistant that can help users move from “I need this product” to a fully informed, user-authorized checkout, all through a natural conversation.

DealPilot Agent is our exploration of what happens when autonomous AI, Android automation, and PayPal-powered commerce come together.