Why I built it
Most support tickets are the same few requests: where’s my order, can I get a refund, how do I change my plan. Basic chatbots can only point people to an FAQ. I wanted to see how far an agent could safely go if it could look things up and take real actions, within strict rules.
What I built
Answers from the docs
Replies in plain language using the store’s help centre, and links to the page it used.
Looks up live data
Checks orders, accounts and deliveries through the store’s own systems.
Takes action safely
Refunds, cancellations and changes happen only within set limits.
Knows its limits
Anything outside the rules goes to a human agent, with the full conversation attached.
Full audit trail
Every lookup and action is logged, so it’s clear exactly what it did and why.
Works where customers are
It runs as a chat widget on the website, and the same agent can answer email and Slack.
How it works
- Customer asksThey ask in their own words, in the chat on the site.
- Agent plansIt decides which tools it needs: an order lookup, a policy check, a refund.
- Rules are checkedA policy layer approves every action before any money moves.
- Done and loggedThe customer gets an answer in seconds; the support team gets a clear record.
Built with
- Python
- FastAPI
- LLM tool calling
- RAG
- PostgreSQL
- Redis