AI Date Planner
AI agent that plans dates in Singapore using live weather, Google Maps and web search, and remembers each user's preferences and date ideas
Problem
Planning a date means juggling the weather, venues and half-remembered ideas across different apps. The agent does it in one conversation, grounded in live Singapore data, and remembers what each person likes.
My role
Built the backend: the agent, its tools, long-term memory, tests and evaluation.
What I built
- Grounded agent: ReAct-style agent that grounds suggestions in live data instead of making them up: Singapore government (NEA) weather, haze and UV readings, Google Places for real venues, and Tavily web search. It picks the forecast that matches how far away the date is (2-hour, 24-hour or 4-day).
- Long-term memory: Trustcall extracts a structured user profile (budget, interests, dietary restrictions, partner) and a date-idea list (idea → planned → booked → completed) from the conversation and saves both to Supabase Postgres, so the frontend can read them directly.
- Durable sessions: Conversation state saved with LangGraph checkpointers in Postgres, so sessions survive restarts in production, plus a Docker Compose stack (Redis + Postgres) for testing in a prod-like setup.
- Testing & evaluation: pytest suite for routing and every tool with mocked HTTP (respx). LangSmith tracing, with LLM-as-a-judge evaluation that scores tool selection and memory extraction separately from final-answer quality.
Stack
- Python
- LangGraph
- LangChain
- OpenAI GPT-4o
- Trustcall
- Postgres
- Supabase
- LangSmith
- Docker
- pytest
Screenshots
