Neural Terp

Inspiration

University life is overwhelming — juggling deadlines, advising appointments, course registrations, and mental health, all across a dozen different portals. We wanted to build something that felt less like a search engine and more like a knowledgeable friend who already knows your schedule, your courses, and your goals. The rise of agentic AI made us ask: what if your assistant didn't just answer questions, but actually did things on your behalf?

What it does

Neural Terp is an agentic AI companion built specifically for UMD students. You can ask it anything from "what assignments are due this week?" to "whats the closest library" — and it acts. It pulls your real course schedule and deadlines from Canvas, routes sensitive mental health queries to a dedicated wellbeing agent, classifies intent to decide when to act vs. when to just inform, and maintains memory of your preferences across sessions. Everything is grounded in UMD-specific knowledge via TerpAI integration.

How we built it

The backend is a FastAPI service built around an orchestration pipeline: an intent classifier routes each message through a safety layer (FERPA-aware), a context builder, and a knowledge router that decides between TerpAI (UMD's own AI, automated via Playwright) and OpenAI as the reasoning engine. Domain-specific sub-agents handle academic, wellbeing, and career queries. Canvas data is scraped with Playwright using UMD's CAS + Duo authentication flow. The tool registry enforces per-user permissions, rate limiting, and audit logging on every action.

Challenges we ran into

Getting Playwright to reliably handle UMD's multi-step authentication — CAS login, Duo push approval, and the optional "trusted device" prompt — required careful polling logic and graceful fallbacks. FERPA compliance added real constraints: we had to design the safety and audit layers before writing any feature code, not after. We also hit a wall trying to integrate Google Workspace for email and calendar because UMD's Google org restricts API project creation — forcing us to rethink that feature entirely.

Accomplishments that we're proud of

We built a working agentic pipeline that goes from a student's natural language message all the way to a real executed action — booking a counseling appointment, fetching live Canvas deadlines, or surfacing course context — with FERPA-safe audit logging on every step. The intent classification, multi-agent routing, and memory system all working together felt like a genuine leap from chatbot to actual agent.

What we learned

Agentic AI in a regulated environment (FERPA, university IT policies) is a fundamentally different problem from building a general chatbot. Safety and compliance have to be first-class citizens in the architecture, not bolt-ons. We also learned that browser automation is surprisingly viable as a real integration layer when official APIs are unavailable — Playwright closed doors that OAuth couldn't open.

What's next for Neural Terp

We want to replace the remaining stub adapters with live integrations — starting with the UMD Student Information System and expanding Canvas scraping into grade and syllabus parsing. A mailto and .ics-based email and calendar feature is next on the list, requiring zero OAuth. Longer term, we're exploring a student-facing browser extension and a voice interface, and we'd love to work with UMD OIT to get official API access so Neural Terp can become a sanctioned campus tool.

Built With

• celery
• cloudforce
• css
• fastapi
• javascript
• jinja
• langchain
• langsmith
• openai
• playwright
• python
• redis
• sqlalchemy
• sqlite
• tailwind