SpendPilot AI: AI-Powered Shopping Companion

Inspiration

Online shopping can be overwhelming. Finding the right product often means navigating complicated filters, visiting multiple retailer websites, comparing technical specifications, checking prices, and managing separate shopping carts.

We wanted to make this process simpler.

What if shoppers could explain what they need in everyday language and let an AI assistant guide them through product discovery, comparison, and shopping management?

That idea inspired SpendPilot AI, a conversational shopping assistant designed to make online shopping smarter, simpler, and more organized.

What It Does

SpendPilot transforms natural-language shopping requests into a guided product discovery experience.

Instead of filling out long forms or navigating complicated filters, shoppers can simply describe what they want. For example:

"I need a laptop for university and programming within a $1,200 budget."

SpendPilot understands the request, gathers relevant preferences through natural conversation, and helps shoppers discover and compare suitable products.

Key Features

Conversational AI Shopping: Shoppers can describe their needs naturally, while the AI understands product categories, budgets, preferences, and requirements.
Intelligent Clarification: Asks relevant follow-up questions to understand the shopper's needs before starting a product search.
Live Product Discovery: Uses Channel3 to retrieve product listings, retailer offers, available pricing, and direct merchant links.
Product Comparison: Organizes product specifications into side-by-side comparisons, helping shoppers understand important differences and trade-offs.
Product Configuration: Provides a visual interface for exploring available options such as storage, colors, and other configurations when supported by the retrieved retailer data.
Multi-Product Shopping Workspace: Lets shoppers manage multiple shopping tasks, select products, and organize their choices in one shopping list.
PayPal Multi-Seller Sandbox Demonstration: Integrates PayPal Orders v2 to demonstrate a multi-seller checkout flow with separate purchase units and controlled Sandbox business sellers.

How We Built It

SpendPilot is built using Next.js, React, TypeScript, and Tailwind CSS. We designed the application around three main parts: conversational AI, product discovery, and payment integration.

1. Google Gemini: Conversational Intelligence

We used the official @google/genai SDK to power SpendPilot's conversational shopping experience.

Gemini helps interpret natural-language requests, identify product requirements, understand budgets and preferences, ask relevant clarification questions, and support product comparisons.

The application also uses a structured shopping workflow to keep track of requirements and maintain context across different shopping tasks.

2. Channel3: Product Discovery

We integrated the @channel3/sdk to retrieve product information and retailer offers from external commerce catalogs.

Channel3 helps SpendPilot discover products across merchants and retrieve information such as prices, product specifications, availability data, and direct retailer links.

We then normalize the retrieved information into a consistent structure so shoppers can compare products more easily.

3. PayPal: Multi-Seller Checkout Demonstration

We integrated the PayPal JavaScript SDK and PayPal Orders v2 REST API to demonstrate a multi-seller checkout workflow.

The demonstration creates an order containing separate purchase units assigned to two controlled Sandbox business sellers. It is designed to demonstrate one buyer approval followed by separate seller payment captures.

Important distinction: The PayPal Sandbox demonstration uses a fixed catalog of sample SpendPilot products and controlled demo sellers. It is separate from the real retailer offers retrieved through Channel3. Channel3 retailer links currently direct shoppers to the respective merchant websites. The Sandbox integration does not process real payments or send payments to arbitrary Channel3 retailers.

4. Application Architecture

We use Next.js server-side API routes to connect the frontend with Gemini, Channel3, and PayPal.

Sensitive API credentials remain on the server. We also use structured validation for shopping requests, budgets, product eligibility, shopping state, and payment orders.

The application is deployed on Vercel.

Challenges We Faced

Building SpendPilot involved several challenges beyond simply connecting an AI model to a chat interface.

Making AI shopping reliable: An AI assistant should not invent products, specifications, or prices. We needed to connect conversational reasoning with external product data and apply validation before presenting results.

Inconsistent retailer information: Different retailers describe the same product in different ways. Normalizing specifications and presenting comparable information required additional processing.

Maintaining conversation context: Shoppers may change their requirements or start completely different shopping tasks. We needed to keep each task separate so information from an earlier search would not affect a new one.

Product configuration accuracy: Retailers can offer different storage options, colors, and configurations. Showing options reliably requires actual configuration data rather than assumptions based on a product title.

Multi-seller payment integration: Working with PayPal Orders v2 introduced additional complexity around seller identification, order validation, purchase units, Sandbox credentials, and separate payment captures.

These challenges helped us understand that building a useful AI shopping assistant requires careful coordination between AI reasoning, external data, application logic, and payment services.

What We Learned

Building SpendPilot taught us that an effective AI shopping experience requires much more than a chatbot.

We learned how to:

Integrate Google Gemini through the official SDK.
Build real-time conversational interfaces with streaming responses.
Connect external product catalogs using Channel3.
Normalize and compare product specifications.
Maintain structured shopping context across multiple tasks.
Apply server-side validation to product discovery and checkout workflows.
Integrate PayPal Orders v2 and test multi-seller payments using Sandbox accounts.
Deploy and maintain a full-stack application using Vercel and GitHub.

Most importantly, we learned how different technologies can work together to turn a simple shopping conversation into a more complete product discovery experience.

What's Next

SpendPilot is an evolving project, and there is still room to improve.

Our next steps include improving retailer configuration verification, expanding product coverage, strengthening price and specification accuracy, and exploring a complete marketplace architecture with proper retailer onboarding and payment authorization.

Our long-term vision is to make online shopping more conversational, transparent, and convenient.

SpendPilot AI: Less searching. Smarter shopping.