Phia Gifting — AI-Powered Gift Discovery

Describe the person, not the product. AI finds gifts they'll actually love.

Comment

Phia Gifting — AI-Powered Gift Discovery

Why this problem?

A few weeks back I spent hours trying to find a birthday gift for a friend who's into yoga and loves Lululemon. Every site I tried just showed me generic yoga mats or the same Lululemon gift card. None of them understood that she's a specific person with specific taste. I ended up buying a gift card — and it felt lazy, like I hadn't put any thought into it.

But this isn't just my problem. Everyone faces it. We've all been there — scrolling endlessly, giving up, and defaulting to a gift card because nothing feels right. Gift cards say "I couldn't figure out what you'd like." A great gift says "I know you."

The core issue: every recommendation system today optimizes for the shopper's own preferences. But gift-giving is fundamentally different — you're trying to understand someone else's personality, interests, and tastes. Phia already solves the "find the best deal" problem brilliantly. We wanted to solve the step before that: figuring out what to buy in the first place.

The Solution

Phia Gifting flips the shopping model: instead of searching for products, you describe the person. Our AI understands who they are — their age, interests, personality, the occasion, your relationship — and searches the internet to find gifts matched specifically to them.

Key features:

  • Natural language input — describe the recipient like you'd tell a friend: "My sister is 28, loves hiking and photography, very outdoorsy"
  • Smart follow-up questions — if your description is vague, the AI asks 1-2 clarifying questions in a chat interface before searching
  • Live product search — searches Google Shopping across Amazon, Walmart, Target, Etsy, and thousands of stores in real-time
  • Diversity-enforced results — if someone likes cricket AND Harry Potter AND photography, you get gifts for ALL three interests, not 9 cricket jerseys
  • Confidence-weighted ranking — uses Wilson score statistical lower bound for review confidence, not just star ratings. A 4.5★ product with 5,000 reviews ranks higher than a 5.0★ with 3 reviews
  • "Why This Gift?" explanations — every recommendation includes an AI-generated explanation of why it fits the recipient
  • Compare Prices — inline price comparison across stores for every product
  • Wishlist — save gifts for later, persists across sessions
  • Refine without restarting — adjust budget or add details from the results page

Technical Stack

  • Frontend: React 18, TypeScript, Vite, Tailwind CSS
  • Backend: Python, FastAPI, Pydantic
  • AI: Claude Sonnet (Anthropic API) — 3 calls per search: recipient parsing, query generation, scoring & explanations
  • Product Search: SerpAPI (Google Shopping) — 8 parallel searches with price range filters and diversity enforcement
  • Scoring Algorithm: Wilson score lower bound + tag matching + occasion matching + gift quality tiers
  • Deployment: Render (backend web service + frontend static site)
  • Storage: localStorage for wishlist, sessionStorage for result persistence

Technical Challenges

1. Diversity problem — Early versions returned 12 Messi jerseys when someone liked "football and Harry Potter." We solved this by enforcing a max of 4 products per category BEFORE Claude sees the candidates, forcing diverse selection.

2. Claude returning non-JSON — Claude occasionally adds preamble text before JSON responses. We built a robust JSON extractor that bracket-matches to find JSON objects/arrays regardless of surrounding text, with separate handling for expected objects vs arrays.

3. Budget respect — Google Shopping defaults to showing cheap popular items regardless of budget. We added the tbs price range parameter to SerpAPI queries AND budget-aware search terms ("premium", "luxury" for high budgets) to surface appropriate products.

4. Store URLs — SerpAPI shopping results only provide Google redirect URLs, not direct store links. We built a store-specific URL builder that generates direct search URLs for 12+ major retailers (Amazon, Walmart, Target, Etsy, etc.) with simplified product names for better search accuracy.

5. Follow-up intelligence — Getting the right balance of when to ask vs when to search directly. Too aggressive = annoying. Too passive = bad results. We tuned it to only ask when age or interests are genuinely missing, and skip when the description is rich enough.

Links

Note: The backend runs on Render's free tier — first request may take ~50 seconds to cold start after inactivity.

Built With

Share this project:

Updates