DingDing Shufang AI platform
Since January 2026 I have been the AI engineer at the headquarters of a large education group, designing and building its business platform and the agents that work on it. I am the main author of the repositories, with a colleague handling the business side. The product is internal, so this page describes the architecture and nothing else.
The problem
The group runs enrollment, payments, CRM and teaching across many cities. The goal was one platform where AI can take real work off staff without ever being allowed to change business facts unchecked.
How it works
- A Go API is the single source of truth for permissions, approvals, audit, payments, CRM, enrollment and teaching, with 30+ background workers and operations commands.
- The React front end, the parent and teacher WeChat Mini Programs, a license service and a question bank each ship independently and consume the API contracts.
- The AI orchestration layer manages model providers, retrieval and business agents. It is kept apart from business facts, and every write goes through an approval contract.
- The agent runtime is being rewritten in Rust, with the existing Python runtime kept as the parity baseline.
- Huoke is a Xiaohongshu operations workspace: staff ask in Feishu and get evidence-backed reports, positioning advice and drafts, all traceable. Compliance checks run before generation, paid model calls have cost reservation and caps, and the Windows and macOS clients ship with install checks and auto-update.
- An Ed25519-signed license service issues activations for the desktop clients.
Built with
- Go
- PostgreSQL
- React
- Rust
- Python
- Temporal
- MCP
- WeChat Mini Programs