library study space demo
dining hall options for high protein meals
landing

OpenCampus

Inspiration

Columbia students constantly juggle multiple websites to check basic campus info — gym hours, dining menus, study room availability, library books. There is no single place that answers these questions quickly.

OpenCampus solves this by providing a conversational assistant that delivers real-time campus data in plain language.

What it does

OpenCampus is a chatbot that gives Columbia students instant answers about four core campus services (for now):

Gym
Real-time hours for Dodge Fitness Center, Uris Pool, Blue Gym, Levien, Aerobics Room, and Functional Fitness Studio, including live activity schedules.
Dining
Daily menus across JJ's Place, Ferris Booth, and John Jay with dietary tags and nutritional macros (protein, carbs, fat, calories).
Study Rooms
Live availability across 29 rooms in 5 libraries with specific time slots.
Library Books
Search Columbia's CLIO catalog for availability, location, call number, and copy count.

Live Demo — Real Data (April 12, 2026)

Every response is grounded in live data scraped from Columbia systems.

Source	Data Point	Method
Gym	6 facilities with activity schedules (e.g. Basketball 10am–1:45pm)	HTML tables + Google Calendar iCal
Dining	167 menu items across 10 meal periods with macros	Angular scope extraction (Playwright)
Study Rooms	302 available slots across 29 rooms, ~100ms latency	Direct LibCal API (HTTP POST)
Library	23 books with copy-level metadata	CLIO Blacklight JSON API

Example query:

"Is there a basketball game at Dodge right now?"
Returns a real-time answer from Google Calendar data.

How we built it

Backend (Python)

Each data source uses a purpose-built scraping strategy:

Dining
Playwright loads the Angular site and extracts structured data via page.evaluate().
Gym
HTML parsing for hours + iCal feeds for activity schedules using icalendar and recurring-ical-events.
Study Rooms
Direct POST to /spaces/availability/grid endpoint (~100ms latency).
Library
CLIO Blacklight JSON API via .json endpoints.

Nutritional macros are batch-estimated using Claude Haiku and cached for near-zero marginal cost.

Frontend

Next.js chat interface (iPhone-style UI)
Scraped data injected into LLM context for grounded responses
Designed for future iMessage integration (SendBlue pending)

Challenges

macOS com.apple.provenance blocked git, pip, and pytest workflows
Each source required a different extraction strategy (Angular, HTML, iCal, REST, Blacklight)
LibCal API required reverse-engineering (eid=-1, exclusive end-date) via XHR inspection
pytest-asyncio version conflicts caused test hangs

Accomplishments

100% real data — no mock responses
~100ms study room queries with no caching
78 menu items with cached macro estimates
22 textbooks pre-indexed with detailed metadata
3/4 sources require zero API keys

What we learned

XHR inspection is more reliable than static analysis
Public iCal feeds are high-value, underused data sources
LLM-based macro estimation is accurate enough for real use
Blacklight exposes a hidden JSON API via .json endpoints

What's next

Raspberry Pi deployment
On-campus node with cron jobs (15-min refresh for gym/dining, real-time for others)
OpenClaw intent parsing
Replace keyword matching with LLM-based multi-turn understanding
Expanded calendar integration
Classes, campus events, office hours via iCal feeds
Proactive notifications
Alerts for study room openings and book availability
Playwright on ARM
System Chromium on Pi for browser-based scraping