Inspiration
What happens when the cheapest product is the wrong product?
That question inspired me to build ProcurePilot AI.
Businesses purchase equipment, supplies, and other products every day. But finding the lowest price doesn't always mean making the best decision. A cheaper product might have the wrong specifications, be unavailable, or fail to meet the company's requirements.
Imagine a business needing two 27-inch office monitors. A cheaper 24-inch monitor might look attractive, but it doesn't meet the requirement.
I wanted to build something that goes beyond recommending cheap products. I wanted a system that helps businesses choose correctly, understand why, approve confidently, and keep a record of every decision.
That's how ProcurePilot AI was born.
What it does
ProcurePilot AI is an intelligent purchasing assistant that helps businesses find suitable products, make informed decisions, approve purchases, and track payments.
Here's how it works:
- Create a purchase request: Describe what you need, specify the quantity, requirements, and budget.
- Discover real products: Using Channel3, ProcurePilot searches actual product listings and compares available offers.
- Check product suitability: The system identifies products that don't meet mandatory requirements, even when they're cheaper.
- Get AI-powered analysis: Google Gemini helps explain the available options and supports better purchasing decisions.
- Approve the purchase: A human reviews and authorizes the selected purchase order before payment.
- Pay through PayPal Sandbox: The application demonstrates a real PayPal API checkout and payment-capture workflow using test accounts.
- Generate a Decision Passport: A detailed record explains what was selected, why alternatives were rejected, who approved the purchase, and the test-payment details.
The Decision Passport
The Decision Passport is one of ProcurePilot AI's defining features.
Instead of ending the process with a simple payment confirmation, ProcurePilot preserves the evidence behind the purchasing decision.
It brings together product comparisons, rejected alternatives, approval history, and PayPal Sandbox payment information in one record that can also be exported.
The goal is simple: every purchase should have a reason.
How I built it
I developed ProcurePilot AI as a solo project, combining modern web technologies, artificial intelligence, product discovery, and PayPal's developer platform.
Frontend: I used React, JavaScript (JSX), and Vite to build a responsive interface, including the purchasing dashboard, product comparisons, approvals, and order management.
Backend: I built the backend with Python and FastAPI to handle business logic, product validation, authentication, purchase orders, and payment operations.
Database: I used PostgreSQL, hosted through Neon, to store purchasing requests, product evidence, approvals, transactions, and audit records. SQLAlchemy and Alembic manage database access and migrations.
Artificial intelligence: I integrated Google Gemini through Google's Agent Development Kit (ADK) to provide explanations based on product evidence. I kept important purchasing rules in Python rather than leaving those decisions to AI.
Product discovery: I integrated Channel3 to retrieve real catalogue listings, product information, retailer offers, and pricing observations.
Payments: I integrated PayPal REST APIs and Orders v2 Sandbox to support test checkout, buyer authorization, and payment-capture verification.
Deployment: I deployed the application using Vercel.
I deliberately designed the architecture to separate AI recommendations, purchasing rules, human approval, and payment confirmation.
AI explains. Code validates. Humans approve. PayPal confirms.
Challenges I ran into
Building ProcurePilot AI involved several challenges.
1. Making AI recommendations trustworthy
AI can generate convincing answers, but purchasing decisions require reliable facts. I had to ensure that AI couldn't override mandatory requirements or invent product prices.
I addressed this by separating AI analysis from the backend rules that determine whether a product qualifies.
2. Handling incomplete product information
Real product listings don't always contain every specification. I had to design checks that could recognize unsuitable or insufficiently verified offers rather than simply recommending them.
3. Integrating PayPal correctly
Implementing PayPal Sandbox required handling separate buyer and merchant accounts, payment authorization, transaction verification, and capture.
One important lesson was that returning from a payment page doesn't automatically mean a payment has been completed.
4. Managing external service limitations
API rate limits, occasional service failures, and deployment restrictions made reliability challenging. I introduced bounded requests, fallback behavior, error handling, and clearer loading feedback.
5. Building everything independently
As a solo developer, I handled the frontend, backend, database, AI integration, payment workflow, testing, and deployment.
Balancing all those components while keeping the application understandable was one of the biggest challenges.
Accomplishments that I'm proud of
I'm particularly proud of building an integrated prototype that goes beyond a typical AI product-recommendation chatbot.
My main accomplishments include:
- Integrating real product discovery through Channel3.
- Connecting Gemini through Google ADK for evidence-based analysis.
- Implementing rules that reject products violating purchasing requirements.
- Building a human-controlled purchasing approval workflow.
- Integrating PayPal Sandbox checkout and payment verification.
- Successfully completing a PayPal Sandbox test transaction.
- Creating the Decision Passport to preserve purchasing evidence, approvals, and payment records.
- Developing the full application independently, from interface design to backend architecture.
The achievement I'm most proud of is the Decision Passport because it turns a simple product recommendation into a decision that can be reviewed and explained.
What I learned
Building ProcurePilot AI taught me that powerful AI isn't just about generating intelligent answers.
It's also about knowing when AI should not make the final decision.
I learned how important it is to separate recommendations from business rules, keep humans responsible for financial approvals, and verify external payment results rather than trusting what appears on the screen.
I also gained practical experience integrating multiple services into one application, designing reliable backend workflows, handling imperfect external data, and making complex processes easier for users to understand.
Most importantly, I learned that the best purchasing decision isn't necessarily the cheapest one. It's the one that actually meets the need.
What's next for ProcurePilot AI
ProcurePilot AI currently demonstrates an accountable purchasing workflow using live product discovery, AI assistance, human approvals, and PayPal Sandbox payments.
My next goal is to develop it into a more complete procurement platform for businesses.
Planned improvements include:
- Real retailer ordering: Connect approved purchase orders directly to supported merchants and suppliers.
- Business workspaces: Introduce organization-level accounts, purchasing teams, and configurable approval policies.
- Smarter comparisons: Expand product discovery, supplier evaluation, and purchasing insights.
- Budget management: Help organizations monitor spending and purchasing limits.
- Order fulfillment tracking: Follow actual supplier orders from confirmation to delivery.
- Business integrations: Connect procurement records with inventory and accounting systems.
- Real-world testing: Work with businesses to validate usability and measure purchasing outcomes.
The current PayPal integration uses Sandbox payments, so it does not send real money to catalogue retailers or place real supplier orders. Those capabilities are part of my future roadmap.
My vision is to make ProcurePilot AI a trusted purchasing assistant that helps businesses spend more wisely, reduce avoidable mistakes, and understand every purchasing decision.
ProcurePilot AI — Every Purchase Has a Reason.
Built With
- adk
- alembic
- channel3
- fastapi
- gemini
- javascript
- jsx
- jwt
- neon
- paypal
- postgresql
- pytest
- python
- react
- restapi
- sqlalchemy
- vercel
- vite
- vitest