TableUs

AI-assisted restaurant discovery for eating out with friends—without the group chat spiral over “where should we go?”

---

The problem

Planning a shared meal is stressful: different budgets, different taste profiles, dietary quirks, and no one wants to be the decider. People also miss out on what’s actually good nearby—the spots that reflect the neighborhood they’re in, not just a generic chain list.

TableUs is built to reduce that friction: it learns how each person likes to eat, blends preferences when you plan as a group, and grounds recommendations in a real place (coordinates + live venue data) so suggestions feel local, cultural, and explorable, not random.

---

Inspiration

The product takes cues from social, taste-forward dining apps: natural language as the main interface, a visual “orbit” of nearby spots, and a friends layer so group plans aren’t a one-size-fits-all ranking. The goal is a hackathon-grade demo that feels like a coherent product: profile → friends → location → search → ranked, explainable picks.

---

Workflow

What people do in the app

1. Pick who you are — Demo accounts in the sidebar; each has a seeded taste profile and reviews.
2. Friends — Add or remove friends on a mutual graph; optional blend of group preferences for a quick “group taste” snapshot.
3. Discover — Set current location (geocode or saved place). The app loads nearby restaurants into the orbit; you type what you want in plain language and optionally @ mention friends for a group search.
4. Review — Drop a natural-language review; the backend updates that user’s taste profile with Gemini.
5. Profile — See how preferences are represented for the active user.

What happens under the hood (search pipeline)

Roughly:

1. Google Maps Platform — Geocoding turns a place name into lat/lng; Places Nearby Search builds a real candidate pool (names, ratings, price, photos, addresses, geometry) around that point.
2. Google Gemini — Interprets the query and each diner’s textual taste profile:
   • Solo search: infer cuisines / intent, then rank and explain why top venues fit this user (and distance context).
   • Group search: merge multiple profiles into one group preference summary, then rank for the combined group with short reasoning.
3. Fallbacks — Gemini calls use a model cascade (e.g. lighter flash models first, then stronger ones) when quotas or availability differ; restaurant cards in the UI stay tied to Places-backed fields where available.

So: Maps = “what exists here”, Gemini = “what matches us, in words we understand.”

---

Core mechanics

Mechanic	Role
Food photo analysis	Upload a dish image; Gemini Vision returns dish, cuisine, and flavor-oriented description.
NL reviews → taste profile	Reviews feel like texting; Gemini rewrites the stored bullet-style profile.
Location-aware search	Nearby set from Places; Gemini ranks against the user’s profile and query.
Group search	Selected / mentioned friends → merged preferences → group-aware ranking and summary.
Friends graph	In-memory mutual add/remove; supports group search and social UX on Discover.

---

Tech stack

Layer	Choice
Frontend	Next.js 16, React 19, Tailwind CSS, Framer Motion
Backend	Python, FastAPI
AI	Google Gemini (vision + text; JSON-structured prompts for cuisine detection, ranking, merging)
Places / maps	Google Maps Geocoding API, Places Nearby Search (and related helpers for photos/metadata)
Data (demo)	In-memory users, friends, reviews—no database required for the hackathon build

Configuration: place GEMINI_API_KEY and GOOGLE_MAPS_API_KEY in backend/.env so Gemini and the Places/geocoding pipeline can run against live APIs.

---

Demo users

User	Sketch
Sam Kwak	Japanese + Italian, polished dinners, umami / savory
Bob Martinez	Mexican + Thai, casual, spicy / bold
Carol Washington	Ethiopian + Indian + Korean, communal, complex spices
William Kang	Korean + Japanese comfort food, lively group spots
Maya Patel	Mediterranean + Indian, bright shareable plates
Nina Okonkwo	New to the graph—shows up on Add friends until someone adds her; light seeded profile + review

---

API surface (reference)

Method	Path	Description
POST	/api/food/analyze	Food photo → Gemini Vision analysis
POST	/api/reviews/submit	Submit review → persist → update taste profile (Gemini)
GET	/api/reviews/{user_id}	List reviews for a user
GET	/api/profile/{user_id}/taste	Taste profile text (+ structured parse when available)
POST	/api/preferences/blend	Blend taste text for a group of user IDs (Gemini)
POST	/api/location/resolve	Geocode a free-text location (Google)
POST	/api/restaurants/nearby	Nearby restaurant pool (Places) for Discover
POST	/api/restaurants/search	NL search + Gemini ranking (single user)
POST	/api/restaurants/search-group	Merged group preferences + Gemini ranking
GET	/api/friends/{user_id}	Friend list
POST	/api/friends/add	Mutual add
POST	/api/friends/remove	Mutual remove
GET	/api/users	List demo users
GET	/api/users/{user_id}	Single user
GET	/health	Health + flags for Gemini / Maps configuration

---

TableUs — less stress picking the table, more time enjoying the meal and the place you’re in.