Inspiration
Inspired by what I've experienced through using multiple AI agents for shopping, I want to design an agent that understands what you need, suggests the best options, and takes you through the whole purchase.
What it does
Describe what you need and your budget. An LLM works out which products would suit what you said, searches for them, and recommends three with a reason for each. It can also put together a set of items within one budget. It never spends money by itself: every payment and refund needs your click, and payments go through the PayPal sandbox.
How we built it
Python and FastAPI on the back end, a single HTML page on the front. Products come from Channel3, the LLM is DeepSeek, and payments use PayPal's Orders and Payments APIs in the sandbox. The code was written with Claude Code as an AI coding assistant, without an agent framework. The model can search and suggest, but creating, capturing and refunding a payment only happen after the user clicks.
Challenges we ran into
Testing with the real model showed it didn't always follow the prompt: asked for "a housewarming gift for a friend, under $40", it asked what kind of gift instead of searching. I added explicit rules and examples to the prompt and kept the checks that really matter in code.
Accomplishments that we're proud of
What I'm proud of is the LLM ranker for each shopping recommendation. For a request like "a smart mug, a notebook and a candle under $200", the LLM ranks the options for each item on its own, and code then tries every combination and picks the one with the best ranks whose total fits the budget. If even the cheapest set is over budget, Pickwise tells you that price instead of guessing, and a chosen set is paid as one PayPal order with a line item per product.
What we learned
LLMs are good at turning a vague request into concrete things to search for, but budgets and prices are safer checked in code than in a prompt.
What's next for Pickwise
The next for Pickwise is to collect real customer reviews and build a RAG layer over them. Index the reviews of each candidate product, retrieve the passages that match what the user asked for and pull out the pros and cons buyers mention. The candidates selected will then be ranked by an LLM ranker and recommend to the users.
Built With
- channel3
- claude
- deepseek
- impeccable
- javascript
- python
- render
