~/projects$ cat movicar/README.md
MOVICAR: multi-agent assistant
Chat assistant that quotes repairs and books appointments for a car repair shop, built with LangChain and LangGraph.

## README.md
The shop gets requests over WhatsApp and phone, and a single person handles them and writes each appointment down in three places. The assistant talks to the customer, asks only for missing data, quotes with the real price list and proposes slots with the right technician.
An orchestrator agent delegates to a quoting agent and a scheduling agent. The design rule is that AI decides what to do and when, while code guarantees how: no agent writes to the database, every write goes through a deterministic layer that validates and applies business rules.
It started as a 61-node n8n flow, moved to a 19-node supervisor pattern, and ended up rebuilt outside n8n with LangChain, LangGraph and Gemini, with its own interface and Docker Compose.
## Architecture
## Highlights
- Supervisor pattern: orchestrator, quoting and scheduling
- Business rules in code, not in the prompt
- Escalation to the shop manager for risky cases
- Per-session memory with the LangGraph checkpointer
- 47 automated tests and 14 scenarios