AI Automation · 2025
AI Customer Operations — Auto Body Shop
A working private prototype for an auto body shop that combined WhatsApp intake, multimodal damage analysis, Kommo CRM state, and draft repair estimates, with every estimate routed to a human for review and approval.
The Challenge
An auto body shop needed a consistent way to collect vehicle details and mixed media, interpret damage against workshop-specific pricing rules, preserve conversation and quote state, and prepare estimates without removing the owner's final decision.
The Solution
Directed the architecture and AI-assisted implementation of a custom assistant connected to WhatsApp and Kommo. The private prototype guided clients through intake, analyzed images, audio, video, and documents, selected models by task complexity, stored operational state in PostgreSQL, and sent draft estimates to a Next.js review dashboard and PWA. Human reviewers could correct, accept, or reject every quote before it moved forward.
- Agent-generated Python and FastAPI components for message routing, AI orchestration, state management, and CRM synchronization
- Conversation-state routing through intake, assessment, draft quoting, correction, and escalation
- Task-based model routing — lightweight models for simple exchanges and more capable models for damage analysis and tool use
- Multimodal intake — images, audio, video, and documents incorporated into the working context
- Kommo integration — lead, conversation, and pipeline state connected to the operating workflow
- PostgreSQL-backed internal state separated the assistant's working data from the CRM
- WebSocket direct channel to quoting API — submissions and status checks without extra hops
- Next.js dashboard and mobile PWA — mandatory human review, correction, acceptance, or rejection of every draft estimate
The Result
The system reached a working private-prototype stage after approximately six months of development. It demonstrated multimodal intake, draft quoting, CRM updates, and mandatory human review, but the client stopped funding before public launch. It is not currently maintained, and no production reliability or response-time claim is made.