🍽️ BiteTime

Personalized dining recommendations for SJSU students — powered by Bob AI.

🌐 Live site: sjsufood.sagg.in · Built for the IBM AI Challenge @ SpartHill Hacks 2026: AI for Sustainable Impact

Inspiration

Every day, SJSU students walk into The Dining Commons without knowing what to eat. The currently offered menu is a flat list of food with no personal context. We wanted to solve the daily "what should I eat?" friction for our students by building something that learns their taste. Sending them a notification of food they might be interested in on the menu. This way they have a source to know whether they want to spend a meal swipe before actually getting into the cafeteria.

What it does

BiteTime is a personalized dining recommendation web app for SJSU students:

Feature	Description
📅 Live daily menu	Browse breakfast, lunch, and dinner for any date, pulled live from the Dine On Campus API by SpartanEat
👍 Rate dishes	Thumbs up / down on any menu item, saved to your Google-authenticated account
✨ AI recommendations	Unrated dishes you're likely to enjoy are flagged "You might like this" and floated to the top
📧 Daily email digest	Each morning you receive a personalized shortlist of top picks before you even leave for campus

How we built it

Stack

Backend: Django 6.1 · Python 3.12 · SQLite
Frontend: Server-rendered Django templates · Tailwind CSS
AI: Bob Ai · Claude · Codex
ML: fastembed · sentence-transformers/all-MiniLM-L6-v2
Auth: Google OAuth 2.0 (server-side token verification)
Email: Django email backend · scheduled send_recommendations management command
Deployment: Gunicorn · sjsufood.sagg.in

Architecture

Dine On Campus API
      │
      ▼
lib/menu_scraper.py   ←── seed_menu management command (runs per meal period)
      │
      ▼
 MenuEntry / Food  (Django ORM)
      │
      ├──▶  menu view  ──▶  lib/food_recommender.py  ──▶  "You might like this" badges
      │         ▲
      │    UserRating  (thumbs up / down per user)
      │
      └──▶  send_recommendations  ──▶  daily email digest

Recommendation engine

lib/food_recommender.py embeds food names as vectors using all-MiniLM-L6-v2 and computes cosine similarity between a candidate dish and everything a user has liked or disliked:

Exact match short-circuits the model instantly.
Best-like similarity must exceed a threshold (0.55).
Like–dislike margin must be positive enough (0.05) to avoid recommending something that's actually closer to a disliked dish.

This means "Chicken Teriyaki Bowl" gets recommended to someone who liked "Orange Chicken" — purely from dish-name semantics, no ingredient lists required.

Challenges we ran into

No public API docs — The Dine On Campus site is a JavaScript SPA. We reverse-engineered its JSON endpoints, including a multi-step flow (fetch period IDs → fetch menu) and browser-like headers to avoid being rate-limited.
Embedding model cold start — fastembed downloads the model on first use. A failed download or inference timeout must never hide already-saved user votes, so the recommendation step is wrapped in a silent fallback.
Idempotent email delivery — Preventing duplicate emails required a RecommendationDelivery model with a UniqueConstraint on (gmail, recommendation_date) and careful get_or_create + sent_at gating.
Threshold tuning — Finding the right cosine similarity threshold and like-vs-dislike margin to surface genuinely relevant dishes without over-recommending took several iterations.

Accomplishments that we're proud of

🤖 A working AI recommendation pipeline driven entirely by food name embeddings — no nutrition data, no ingredient lists — that still produces intuitive results.
🔄 A robust, idempotent menu scraper that handles API quirks and can be safely re-run for any date and meal period.
🎨 A clean, accessible UI with date navigation, meal period tabs, and per-item vote toggles — server-rendered, zero JavaScript frameworks.
📬 A proactive daily email pipeline with deduplication, error logging, and a full audit trail of every recommendation sent.

What we learned

Semantic text embeddings are a surprisingly powerful signal for food similarity — treating dish names as natural language and measuring their proximity in embedding space works well enough to drive real recommendations.
Reverse-engineering a JS-heavy SPA's internal API (session cookies, multi-step period IDs, CORS headers) takes patience but produces a far more stable integration than HTML scraping.
Building idempotent pipelines (menu seeding, email delivery) from the start saves a lot of pain when things need to be recovered or replayed.

What's next for BiteTime

[ ] Nutrition-aware filtering — incorporate dietary tags (vegetarian, vegan, allergens) to let users set hard preferences
[ ] Community ratings — surface crowd upvote counts so students can see what's popular today
[ ] Push / SMS notifications — replace email with same-morning phone nudges
[ ] Multi-hall support — the architecture is location-agnostic; adding other SJSU locations only requires a new location ID
[ ] Collaborative filtering — blend the content-based (embedding) approach with user–user similarity to surface hidden gems
[ ] Mobile-first redesign — optimise the UI for the phone-in-hand use case students actually have standing in line