Inspiration
International students make up over 1.1 million people studying in the United States — yet when something goes wrong, they're often completely on their own. An OPT rejection letter arrives with no explanation. A landlord threatens eviction using legal language nobody taught them. A W-2 shows up during tax season when they're supposed to be filing as a nonresident alien. Office hours are closed. Friends don't know either. Google returns generic results that don't account for their visa status, their specific university's policies, or the urgency of the moment.
We built CampusGuard AI because we've lived this. As international students ourselves, we know what it feels like to be lost in bureaucracy at 11pm with no one to call.
What It Does
CampusGuard AI is a 24/7 conversational assistant built for international students at Washington University in St. Louis. You describe your situation — or upload a confusing document — and CampusGuard:
- Identifies exactly what's going on and what domain it falls under (immigration, housing, taxes, health)
- Searches the web in real time to find the correct campus office and their contact information
- Generates a professional, personalized email draft and saves it directly to your Gmail inbox
- Walks you through the next steps with clear, plain-English guidance
Two modes power the experience: Campus Navigator (for emergencies and situations) and Document Advisor (for uploaded files like OPT rejection letters, W-2s, lease agreements, and I-20s).
How We Built It
The core AI pipeline runs through Google Gemini 2.5 Flash, which handles natural language understanding, document analysis, and response generation. We use Gemini's built-in Google Search grounding to pull real, up-to-date contact information from .edu websites — so the office email we find is actually correct.
The frontend is built with Next.js 16 and React 19, styled with Tailwind CSS 4. Authentication is handled by NextAuth v4 with Google OAuth, which also captures the Gmail access token needed to create email drafts via the Gmail API.
We built structured knowledge playbooks in Markdown for OPT/STEM extensions, Missouri tenant rights, and international student tax filing — giving the model grounded, authoritative context beyond what a general LLM would know.
Voice input is handled entirely client-side via the Web Speech API — no extra service needed.
Challenges We Faced
Getting Gmail drafts to work reliably was the hardest part. Gemini's tool-calling system is non-deterministic — sometimes it would call the draft tool, sometimes it wouldn't. We had to move away from relying on the model's decision and instead built a deterministic pipeline that always attempts to extract a contact and create a draft when the situation warrants it.
Grounding responses in real data was also tricky. Generic LLM responses are fine for general questions, but international students need to know the actual phone number for WashU's OISS office, not a hallucinated one. We solved this by using Gemini's Search tool and extracting structured contact info from the results before generating the final response.
Scope creep — once we started, we kept finding more situations where students needed help. OPT rejections. Cap-gap. ITIN applications. Security deposits. Emergency mental health. We had to stay disciplined about what to build deeply versus what to acknowledge and hand off.
What We Learned
- How to build reliable agentic pipelines where tool calling needs to be deterministic, not probabilistic
- How to ground LLM responses in live search results to avoid hallucination on high-stakes information
- How to use the Gmail API to create drafts with proper MIME encoding
- That the best product for a specific community is built by members of that community
What's Next for CampusGaurd AI
- Expand beyond WashU to other universities with configurable office directories
- Add support for more document types (I-797 approval notices, DS-2019, CPT letters)
- Build a mobile app with push notifications for deadline reminders (OPT filing windows, tax deadlines)
- Multi-language support — many international students are more comfortable in their native language
Log in or sign up for Devpost to join the conversation.