TerpFlowOS

Flow Transit
Student Home

Inspiration

Campus life is a logistics problem hiding in plain sight: shuttle delays, packed study spaces, safety incidents, ticket exchange risk, and event overload all happen at once.
We wanted to build something that feels essential to student life at UMD — not just a chatbot, but an operational copilot.

What it does

TerpFlowOS is a multi-agent campus logistics platform with a voice copilot.

It includes four core simulation modules:

FlowTransit: adapts to transit delays, outages, and room-capacity constraints
TerpTrade: models safer ticket exchange workflows
TerpFix: triages campus maintenance/safety reports
TerpSync: generates conflict-aware daily recommendations

It also supports:

Student mode with Home, Map, Wallet, and Schedule views
Voice interaction (speech-to-text, reasoning, text-to-speech)
Fallback-aware AI orchestration so the system remains usable under degraded conditions
Moderation + escrow/chat workflow simulation for risk-aware exchanges

How we built it

Frontend: React + TypeScript + Vite + Tailwind + Framer Motion
Backend: FastAPI + Python + SQLAlchemy + JWT auth + WebSocket flows
Data layer: PostgreSQL
AI + Voice:
- TerpAI-first planning
- secondary planner fallback path
- ElevenLabs for speech input/output
Infra: Docker Compose stack, Caddy, Cloudflare tunnel modes for demo access
Testing: Pytest + frontend lint/build checks

We designed the system as modular services so each campus workflow can evolve independently while sharing common auth, routing, and orchestration infrastructure.

Challenges we ran into

Mobile browser voice playback policies (autoplay/permission constraints)
Provider reliability + fallback correctness under missing token/key states
Prompt grounding to avoid incorrect app descriptions
Cross-device demo networking via tunnel/proxy/CORS constraints
Balancing realism vs speed in simulation iteration design

Accomplishments that we're proud of

Built a full-stack, multi-module simulation platform with voice UX
Implemented robust fallback behavior across AI and speech layers
Added practical safety/ops themes (triage, moderation, escrow-like transitions)
Shipped a mobile-aware voice experience with explicit replay fallback
Kept the system runnable locally and demo-able remotely

What we learned

Reliability matters as much as model quality: graceful degradation wins demos and real usage.
Voice UX on mobile requires explicit handling of browser policy constraints.
Multi-agent products need strong contracts between UI, orchestration, and action safety.