AI agents Lab
Customer Service Agent
A tool-calling support agent built with the OpenAI Python SDK, used to practice agent design, guardrails, and evaluation end to end.
- Python
- OpenAI SDK
- gpt-4.1-mini
- Tool calling
- Jupyter
Pipeline
- 01 Customer message Free-text request
- 02 Agent Chooses a tool
- 03 Tools Lookup, policy, actions
- 04 Guarded reply Checked before sending
What it is
A support agent for a simple store, built in a Jupyter notebook with the OpenAI Python SDK and gpt-4.1-mini. It handles common requests by calling tools, such as looking up an order or checking a policy, instead of answering from memory.
What I focused on
- Tool design: small, well-described tools with strict argument schemas, so the model picks the right one.
- Memory: keeping the relevant parts of the conversation without resending everything.
- Guardrails: input checks for off-topic or unsafe requests, and output checks before a reply is shown.
- Evaluation: a fixed set of customer requests with expected tool calls and outcomes, re-run after every change.
Why it’s here
This is the lab where I try techniques before using them in Tech Horizon projects: prompt structure, retrieval, tool routing, and tracing with LangSmith.