Inspiration

I wanted to explore a practical question: can a shopping assistant understand what someone needs without becoming the authority on prices or payments?

For shoppers, describing a need is often easier than searching through a catalog. For small merchants, conversational shopping could make discovery more useful, but recommendations still need to reflect products they actually offer.

That became the idea behind DealPilotAI: let AI help with product discovery, while the application handles pricing and the shopper controls payment approval.

What it does

A shopper can describe what they want in English or Hindi, add a budget, and receive suggestions from a merchant’s catalog. They can select products, change quantities, and review a server-calculated quote before starting PayPal sandbox checkout.

The demonstration uses a simple request: a desk lamp and an organizer within a seventy-five-dollar budget. One of each totals $43.49. Adding a second organizer changes the total to $61.99. Quantity changes require a fresh quote before checkout.

AI selects products and proposes quantities. The server checks those suggestions against the catalog and supplies the displayed prices, currency, availability, and factual explanations. If live inference is unavailable, the interface clearly identifies the fallback.

An authenticated merchant view provides access to order records and payment status.

How we built it

DealPilotAI grew from an earlier Telegram affiliate project. During the hackathon, I added the merchant catalog, conversational shopping experience, cart and quote services, and PayPal integration. The optional affiliate functionality remains separate from merchant checkout.

I kept the application in Python with Flask and Jinja, using JavaScript and CSS for the responsive storefront. Featherless provides the model inference through an OpenAI-compatible client. Pydantic validates structured model responses, and SQLAlchemy handles commerce records with SQLite or PostgreSQL.

PayPal Orders v2 connects the quote to sandbox checkout. The implementation includes approval callbacks, capture verification, repeat-request handling, and recovery paths for uncertain payment outcomes.

I used OpenAI Codex for development and review assistance, and ElevenLabs for the demo narration.

Challenges we ran into

The first challenge was grounding recommendations. Correcting a numeric price was not enough if the model’s explanation could still invent a discount or make an unsupported stock claim. I changed the boundary so that the model proposes selections while the server supplies the commercial facts shown to shoppers.

Payment recovery was another challenge. A timeout does not necessarily mean an order failed, and refreshing a page should not create another payment attempt. I worked on persistent identifiers, ownership checks, transactional updates, and explicit review states for ambiguous outcomes.

Testing across SQLite and PostgreSQL also exposed differences that were easy to miss with one database alone.

Accomplishments that we're proud of

The recorded verification run passed 257 tests, including 16 browser scenarios and three PostgreSQL integration tests. Statement coverage across the AI, commerce, PayPal, and web packages reached 81.69%.

Six genuine AI scenarios passed without fallback, covering English and Hindi requests, budget constraints, multiple products, unsupported requests, and malicious catalog content.

Most of all, I’m proud that the product keeps its responsibilities clear: the model recommends, the server validates, and the shopper decides.

What we learned

A useful shopping assistant needs more than a convincing answer. It needs reliable boundaries between recommendations, catalog facts, cart arithmetic, and payment authorization.

I also learned to distinguish implemented behavior from verified integration evidence. Automated tests help, but they do not replace observing a real external service complete its part of the workflow.

Current demo scope

The video shows genuine AI recommendations and genuine PayPal sandbox order creation. It stops before buyer approval and payment capture. The success popup is explicitly labeled as an illustration, not a paid receipt.

Buyer-approved capture, receipt recovery after a full application restart, and genuine signed webhook delivery remain unverified. This is a sandbox prototype with demo inventory; it does not charge real money or fulfill orders.

What's next for DealPilotAI

My next step is to verify the complete buyer-approved sandbox payment and recovery journey. Beyond the prototype, I would focus on database-backed inventory reservations, operational authentication, monitored background workers, and refund and dispute handling.

Built With

Share this project:

Updates

Submission history