Built with Marcus Cheung, who founded the product while training with me at Preface. My role is technical lead, mentor and reviewer.
The problem
Gift discovery is high friction. The product needed a recommendation pipeline that could explain its picks, real buy links, and a business model, built by a founder I mentor at Preface.
What was built
- A multi-stage pipeline: adaptive questionnaire, discovery through Exa and Serper, Azure OpenAI scoring, then ranking and de-duplication with explainable match notes.
- A signed-in hub with recipient profiles, Discover, Gift Battle comparisons, DIY ideas chat and analytics, gated by Stripe subscriptions through webhooks.
- Vercel cron jobs that populate and de-duplicate the product index, and Sentry plus Upstash rate limiting in front of the AI routes.
Decisions worth explaining
- Firebase holds auth and user data while Supabase holds the global product index with pgvector embeddings, so each store does the job it is good at.
- A reliability plan targets the curated-fallback rate, because a recommendation engine that silently falls back is not doing its job.
Outcome
A commercial product on its own domain with three paid tiers. I lead the architecture and reviews, and the founder ships the features.



