A LangGraph-powered, tool-augmented conversational AI agent designed to automate and optimize rental lead management for real estate agents in Singapore.
Propchat integrates dynamic multi-step reasoning, Excel-based structured data, evaluation pipelines using LangSmith, and containerized deployment for real-world robustness.
Sample Conversation.
Propchat is built using LangGraph to model multi-turn agentic behavior for real estate workflows. The assistant is capable of:
- Structured tool invocation with memory and state-awareness.
- Excel-based data handling to align with local agent practices.
- Autonomous scheduling, lead qualification, and negotiation flow control.
- Observable and evaluable traces using LangSmith for continuous improvement.
It demonstrates advanced agentic patterns with introspectable flow control, model evaluation feedback loops, and a production-ready deployment interface.
realtor-assist/
│
├── Dockerfile # Docker container definition
├── docker-compose.yaml # Multi-service deployment
├── langgraph.json # LangGraph agent definition
├── real_estate_sheet.xlsx # Excel source for listing data
│
├── evaluation/ # Evaluation suite (LangSmith-driven)
│ ├── dataset.py # Generates evaluation dataset from traces
│ ├── helper.py # LangSmith API utilities
│ └── test.py # Prompt/model-level test runner
│
├── agent/
│ ├── agent.py # LangGraph conversational agent
│ ├── requirements.txt # Python dependencies
│ └── utils/
│ ├── nodes.py # LangGraph node definitions
│ ├── states.py # Custom state schema
│ ├── tools.py # Tools used by the agent
│ └── utils.py # Misc utilities
Propchat uses LangGraph to define a state machine representing the assistant's reasoning flow.
Custom tools are defined in agent/utils/tools.py, built with LangChain’s Tool interface and integrated into the graph.
get_agent_listings: Read listing data from.xlsxfile.check_listing_availability: Query availability against structured entries.
These tools interface directly with structured tabular data, supporting workflows familiar to Singaporean real estate agents.
Propchat includes a rigorous evaluation loop using LangSmith:
- Trace-based dataset generation: Evaluates real-world interactions.
- Multi-model prompt comparison: Benchmarks performance across configurations.
- Judged on:
- ✅ Correctness
- ✂️ Conciseness
- 🚫 Hallucination avoidance
Traces are automatically logged to build datasets using evaluation/dataset.py. Evaluations are triggered via test.py.
Using LangSmith’s observability layer:
- Trace every interaction with token-level introspection.
- Compare different runs and prompt versions.
- Identify failure modes (tool misuse, hallucination, broken logic paths, high latency).
git clone https://github.com/yourname/realtor-assist.git
cd realtor-assistCreate a .env file in the root using .env.example:
docker-compose up --buildThis launches the agent and any necessary services. Ideal for local testing, remote deployment, or staging environments.
- ✅ Excel-based structured listing support
- ✅ Agentic LangGraph flows with memory & tools
- ✅ End-to-end LangSmith evaluation suite
- ✅ Dockerized CI/CD pipeline
- 🔜 WhatsApp integration
- 🔜 Auto-sync with calendar availability
- 🔜 Two-way landlord/tenant messaging module
| Feature | Description |
|---|---|
| 🔄 LangGraph Agent | Custom state-machine-based agent with nodes, memory, and transitions. |
| 📈 LangSmith Evaluation | LLM-as-judge metrics: correctness, conciseness, hallucination. |
| 🧩 Excel Tooling | Agents operate directly over .xlsx sheets for local workflows. |
| 📦 Docker Deployment | Easily containerized and deployable with Docker Compose. |
| 📡 Full Observability | Trace logs, prompt diffs, and live monitoring via LangSmith. |
| 🔬 Model Testing | Swap prompts/models to benchmark performance. |
We welcome contributions in tooling, UI integrations (e.g., WhatsApp frontend), or evaluation enhancements. PRs, issues, and prompt optimizations are encouraged.
Built with ❤️ to make every realtor’s day a little smoother.