Inspiration

Every day between lectures, college students face the same dilemma: standing in crowded canteens with short 15-minute breaks, tight budgets, and severe decision fatigue. We wanted to eliminate the "what should I eat?" stress by building a smart campus companion that instantly determines the best meal based on a student's exact budget, cravings, diet, and rush-hour schedule.

What it does

Canteen Cue is an intelligent, campus-grounded food decision assistant for college canteens:

Conversational Decisions: Students can type or speak queries naturally (e.g., "I have ₹80, vegetarian, need something quick in 15 mins").
Strict Menu Grounding: Recommends dishes exclusively from 5 real campus outlets (Anna's Restaurant, Mughlai Affair's, HALKA PHULKA, MAGGIE POINT, and 4TH FLOOR).
Smart Combos & Alternatives: Suggests paired items (e.g., Samosa + Chai), budget trade-offs, and cheaper substitutes when a dish is out of stock or over budget.
Student Auth & Preferences: Supports permanent email/password logins, OTP verification, Google OAuth, and a zero-friction Guest Mode, remembering student taste profiles and chat history across sessions.

How we built it

Frontend Interface: Built with React 19, TypeScript, and Vite, featuring a clean campus-themed UI, responsive cards, and the native Web Speech API for hands-free voice search.
AI & Conversational Layer: Integrated with the Google Gemini API (gemini-flash-latest) for natural language understanding and multi-turn conversational memory, guided by a token-efficient catalog prompt.
Local Deterministic Fallback Engine: A custom TypeScript rule engine (natural language parser, constraint filter, and candidate scoring ranker) that provides instant, offline recommendations with zero token overhead.
Cloud Backend & Database: Supabase (PostgreSQL tables for outlets, 239+ verified dishes, and session syncing) coupled with browser IndexedDB for offline caching.

Challenges we ran into

Eliminating AI Hallucinations: Large language models often invent dishes or misquote prices. We resolved this by building a strict catalog validation layer that forces responses into typed schemas and cross-checks every recommendation against real inventory.
Handling Messy Student Queries: Students rarely speak in structured data. We engineered regex parsers and typo normalizers to handle colloquial terms like "bucks", "ruppes", and "chkn biryani under 80".
Zero-Downtime Hybrid Architecture: Designing seamless fallback routing that automatically switches between cloud Gemini AI and our offline deterministic engine if network latency or API rate limits occur.

Accomplishments that we're proud of

100% Catalog Accuracy: Verified recommendations across 239+ authentic campus dishes without out-of-catalog suggestions.
Sub-Second Response Times: A hybrid router that resolves standard student queries instantly without burning unnecessary API tokens.
Rigorous Verification: Passed 2,000+ automated stress tests validating complex budget edge cases, dietary filters, and multi-turn conversations with zero runtime crashes.

What we learned

How to strictly ground LLMs using compact system prompts and JSON schema enforcement.
How to architect multi-turn conversational state machines that remember student constraints (budget, diet, time limits) over sequential messages.
Designing offline-first web applications using dual storage architectures (Supabase PostgreSQL cloud sync + browser IndexedDB).

What's next for Canteen Cue

Live Canteen Wait Times: Crowdsourced or outlet-updated queue meters so students know preparation delays before walking over.
Direct UPI Ordering & Digital Tokens: Enabling students to order and pay ahead with a digital pickup token to bypass physical lines entirely.
Nutritional & Macro Tracking: Calorie and protein filters to help health-conscious students meet fitness goals on a campus budget.