Thrilled to share that we placed 3rd at HackHarvard 2025 out of over 150 teams.
Julian Ng-Thow-Hing, David Chung, Dheeraj Vislawath, and I built Yumi, an agentic social food network that learns user taste and merges group preferences to help friends agree on where to eat.
The Problem
When we arrived in Boston for HackHarvard, we wanted to explore the city’s food culture but ended up defaulting to a chain restaurant.
Existing apps reduce restaurants to ratings and tags, failing to capture the nuanced preferences people actually have — texture, spice, ambiance, or cultural authenticity. Group decisions devolve into indecision and generic picks. We wanted to build something that truly understands taste and coordinates intelligently.
What Yumi Does
Yumi is an AI-powered coordination platform that:
- Learns each user’s taste from natural language reviews, interactions, and search behavior.
- Lets users type or speak preferences to get personalized restaurant suggestions.
- Builds a taste graph that visualizes friends’ preferences and similarity.
- Merges multiple users’ profiles when friends are tagged with
@mentions, balancing preferences across the group. - Uses a real-time spatial restaurant database powered by Google Places and PostGIS.
- Handles voice-driven restaurant reservations via Twilio and ElevenLabs.
My Contributions (Backend Development)
Led the development of Yumi’s data and AI pipeline:
- Built a backend pipeline that scraped thousands of Boston restaurants via the Google Maps API, using spatial grid partitioning for coverage.
- Designed and implemented a data enrichment workflow that:
- Passed each restaurant through the Gemini API to generate concise, descriptive summaries.
- Used a classification model (trained on Hugging Face) to detect whether uploaded photos contained food or irrelevant content.
- Employed Gemini again to identify cuisine types and dish-level context for food images.
- Structured and stored all restaurant data in Supabase Postgres with PostGIS for efficient geospatial queries and distance-based ranking.
- Collaborated on Gemini function-calling integration, allowing the AI to intelligently retrieve restaurants and merge taste profiles.
Tech Stack
Frontend: SwiftUI (iOS) + Next.js Web Dashboard (React, TailwindCSS)
Backend & AI: Gemini Flash, Google Places API, Supabase/PostGIS, Sentence Transformers, Hugging Face models, ElevenLabs, Twilio Voice API
Deployment: Render (API) + Netlify (Web) + Supabase (Database/Auth)
Challenges
Merging taste preferences across multiple users was nontrivial. Conflicting attributes like spice tolerance and price sensitivity required a dynamic weighting scheme and structured prompt design for Gemini.
Integrating real-world restaurant data with AI-labeled context demanded a reliable scraping and enrichment pipeline that could operate within hackathon constraints.
Outcome
We shipped a working multi-agent system that:
- Automatically builds and updates user taste profiles.
- Accurately labels and categorizes real restaurant data.
- Suggests restaurants that satisfy all members of a group.
- Demonstrates how AI agents can coordinate, learn, and make decisions across multiple users in a real-world setting.
What’s Next
We’re extending Yumi to:
- Suggest meetups automatically based on overlapping preferences.
- Handle reservations via AI voice calls.
- Learn from real dining feedback to refine profiles.
- Expand beyond Boston to new cities with dynamic scraping and labeling.
Tagline
You, Me, and Food — Together we make Yumi.