BU Life AI

A GenAI campus assistant that helps Boston University students navigate campus life — from finding study spots to understanding OPT policies — powered by a multi-agent supervisor architecture.

• Multi-agent system — supervisor routes queries to specialized sub-agents (places, resources, events) using LangGraph
• Hybrid RAG pipeline — BM25 keyword + vector semantic retrieval with RAGAS-evaluated accuracy
• Production-ready — streaming responses, conversational memory, structured output, guardrails, deployed on AWS

---

What It Does

Students ask natural language questions — the agent figures out the intent, calls the right tools, and returns grounded answers:

• "I have 25 mins near CDS, where can I eat?"
• "Find me a quiet study spot near CAS with outlets"
• "How do I apply for OPT as an F-1 student?"
• "Any AI or startup events this week?"

---

Architecture

graph TD
    UI[Next.js Frontend] -->|SSE Stream| API[FastAPI Backend]
    API --> SUP[Supervisor Agent]
    SUP -->|.as_tool| PA[Places Agent]
    SUP -->|.as_tool| RA[Resource Agent]
    SUP -->|.as_tool| EA[Events Agent]
    PA --> DB[(PostgreSQL + pgvector)]
    RA --> RAG[Hybrid RAG]
    EA --> DB
    RAG --> BM[BM25 Retriever]
    RAG --> VEC[Vector Retriever]
    BM --> DB
    VEC --> DB
    SUP -.->|checkpointer| MEM[MemorySaver]
    SUP -.->|store| IMS[InMemoryStore]
    SUP -.->|pre_model_hook| GR[Guardrails]

Loading

---

7 Agentic Features

#	Feature	Implementation	LangGraph API
1	Multi-Agent Supervisor	3 sub-agents converted to tools	create_react_agent + .as_tool()
2	Conversational Memory	Thread-isolated chat history	MemorySaver checkpointer
3	Self-Correcting Loop	Tools return retry guidance on empty results	ReAct loop + smart tool returns
4	Structured Output	Pydantic model for typed responses	response_format=AgentResponse
5	Parallel Tool Execution	LLM calls multiple tools per turn	Native ToolNode
6	Semantic Caching	Cross-thread user preference store	InMemoryStore
7	Input Guardrails	Block prompt injection before LLM	pre_model_hook + trim_messages

---

Tech Stack

Layer	Technology
Frontend	Next.js 14, Tailwind CSS, SSE streaming
Backend	FastAPI, Python 3.12
AI Agent	LangGraph supervisor, LangChain LCEL
LLM	AWS Bedrock — Claude 3 Haiku
Embeddings	AWS Bedrock — Titan Embed Text v1
RAG	EnsembleRetriever (BM25 40% + vector 60%)
Database	PostgreSQL 15, pgvector
Observability	LangSmith tracing (zero code)
CI/CD	GitHub Actions

---

RAG Evaluation (RAGAS)

Ablation study over 20 BU-specific questions:

Experiment	Faithfulness	Context Precision	Context Recall
Baseline (vector, 15 pages)	0.93	0.05	0.07
Expanded corpus (vector, 51 pages)	0.97	0.14	0.13
Hybrid search (BM25+vector, 15 pages)	0.98	0.10	0.12
Both (BM25+vector, 51 pages)	0.93	0.15	0.09

Expanded corpus drove the biggest gain in recall. Hybrid search improved precision for acronym-heavy queries (OPT, CPT, FAFSA).

Run evals: cd backend && python eval/run_eval.py

---

Quick Start

1. Database

psql postgres -c "CREATE DATABASE bulife;"
psql bulife -c "CREATE EXTENSION IF NOT EXISTS vector;"
psql bulife < backend/app/db/schema.sql

2. Backend

cd backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # add AWS credentials
python -m app.db.seed_data
uvicorn app.main:app --reload --port 8000

3. Frontend

cd frontend
npm install
cp .env.local.example .env.local
npm run dev

Open http://localhost:3000

---

Deploy

Component	Platform	Config
Backend	AWS EC2 (Docker)	backend/Dockerfile
Frontend	Netlify	frontend/netlify.toml
Database	Neon PostgreSQL	Connection string in DATABASE_URL
Tracing	LangSmith	Set LANGCHAIN_TRACING_V2=true in env

# Backend on EC2
docker build -t bulife-backend ./backend
docker run -p 8000:8000 --env-file backend/.env bulife-backend