SJSU Safeline

NVIDIA Agents for Impact Hackathon — A compassionate AI agent that helps SJSU students find the right campus or national resource when they're embarrassed or nervous to reach out. No judgment; describe your situation in your own words and get a specific contact plus a message you can send or say.

---

What it does

• Multi-stage reasoning: The agent doesn’t just answer — it assesses urgency, triages your situation, searches the resource database, and drafts a personalized outreach message (email + phone script) using NVIDIA Nemotron.
• SJSU + nationwide: 33 campus resources (CAPS, SJSU Cares, PRIDE Center, ISSS, AEC, etc.) and 13 national hotlines (988, Crisis Text Line, Trevor Project, NDVH, RAINN, SAMHSA, Veterans Crisis Line, and more).
• Visible decision-making: The Agent Trace panel shows the real tool chain (urgency → triage → search → draft) so you can see the agent thinking, planning, and acting.

---

Tech stack

Layer	Stack
LLM / Agent	NVIDIA Nemotron (super-120b orchestrator + nano-30b for tool sub-calls) via langchain-nvidia-ai-endpoints
Backend	FastAPI, LangChain (tool-calling agent), SQLite
Frontend	React, Vite, Tailwind CSS
Streaming	Server-Sent Events (SSE) for real-time response and tool-call events

---

Prerequisites

• Python 3.10+
• Node.js 18+
• NVIDIA API key for NVIDIA NIM / Nemotron

---

Setup

1. Environment

At the project root, create a .env file:

NEMOTRON_API_KEY=your-nvidia-api-key
NEMOTRON_BASE_URL=https://integrate.api.nvidia.com/v1

2. Seed the resource database (one-time)

From the project root:

cd backend/data && python seed.py && cd ../..

This creates backend/data/campus_resources.db with 46 campus and national resources.

3. Backend dependencies

pip install -r requirements.txt

4. Frontend dependencies

cd fe && npm install && cd ..

---

Run

Terminal 1 — Backend (from project root):

uvicorn backend.main:app --reload --host 0.0.0.0 --port 8000

Terminal 2 — Frontend:

cd fe && npm run dev

Open http://localhost:5173. The Vite dev server proxies /chat/ and /resources/ to the backend at port 8000.

---

Agent workflow

1. Assess urgency — Nemotron evaluates the message (immediate / elevated / standard). If immediate, the agent responds with 988/911 and stops.
2. Triage situation — Nemotron identifies resource categories and search terms (e.g. mental health, SJSU Cares, LGBTQ+).
3. Search resources — SQLite lookup by keyword and/or by resource name (e.g. CAPS, 988).
4. Draft outreach — Nemotron generates a short, empathetic email draft and phone script tailored to the student’s situation.

All of this is visible in the Agent Trace panel (top-right) during and after each reply.

---

Project structure

NVIDIA-Agents-for-Impact/
├── .env                    # NEMOTRON_API_KEY, NEMOTRON_BASE_URL (create this)
├── requirements.txt        # Python deps
├── backend/
│   ├── main.py             # FastAPI app, CORS, routers
│   ├── config.py           # Settings (loads .env from repo root)
│   ├── database.py          # SQLite path and get_db
│   ├── agent/
│   │   ├── agent.py        # LangChain tool-calling agent, run_agent(), ToolCallCollector
│   │   └── tools.py        # assess_urgency, triage_situation, search_resources, get_resource_by_name, draft_outreach_message
│   ├── routers/
│   │   ├── chat.py         # POST /chat/ — SSE stream (tool_call, categories, token, done)
│   │   ├── health.py       # GET /health
│   │   └── resources.py    # GET /resources/, GET /resources/{id}
│   └── data/
│       ├── seed.py         # Builds campus_resources.db
│       └── campus_resources.db
└── fe/                     # React + Vite + Tailwind
    ├── src/
    │   ├── App.jsx         # Chat state, SSE handling, agentTrace
    │   ├── api.js          # sendChatMessage (SSE), fetchResources
    │   └── components/     # ChatLayout, ChatMessage, ChatInput, Header, AgentTrace, ResourceCard
    └── package.json

---

License

MIT.