Campus Crew

• 

  Demo Image

Inspiration

Every team runs on documents nobody reads. Policy manuals, brand guides, compliance handbooks;they collect dust until someone breaks a rule nobody remembered. We saw this organizing events at Columbia: four governing docs, four tabs open, zero confidence we weren't violating something. And AI makes it worse; your chatbot will happily write "free beer" on a flyer without knowing alcohol's banned in Section 5.2. The gap isn't intelligence. It's accountability.

What it does

Campus Crew turns governing documents into specialized AI agents. Upload a policy, and it becomes the person who enforces it. Upload a brand guide and it writes on-voice copy. Upload survey data and it speaks for your audience. Agents execute tasks in a pipeline: drafting, reviewing, flagging violations with exact clause citations, and revising until compliant. Upload a new document mid-session, and a new agent spawns, changing the output in real time.

How we built it

React + Pixi.js frontend rendering an isometric campus where agents work visibly in distinct zones. Python/FastAPI backend with SSE streaming. Claude API (Opus) powers document distillation into structured agent profiles and all agent reasoning, with extended thinking for compliance analysis. LangGraph orchestrates the pipeline with conditional routing for the violation-revision loop. ChromaDB handles per-agent document retrieval.

Challenges we ran into

Getting compliance accuracy right. The agent needed to reason contextually ("free drinks" is fine, "free beer" is not) while citing specific clauses. Latency across four sequential API calls required careful prompt optimization to hit our 60-second pipeline target. We also had to brutally cut our original animation scope to ship on time.

Accomplishments that we're proud of

"Free beer" gets flagged every time with a real citation. New documents genuinely change output. Claude is integrated at every layer: extraction, persona construction, RAG reasoning, compliance enforcement, and streaming output.

What we learned

Organizations need AI that reads the manual and won't let violations slide. Grounding and enforcement matter more than raw intelligence.

What's next for Campus Crew

Legal teams reviewing contracts against regulation. HR onboarding against company handbooks. Any team drowning in governing documents. Plus: parallel agent execution, custom roles, and persistent agent memory across sessions.

Built With

• claude