HSK 3 learning agent, from zero

Give the learner a next move, not another word list.

An international Chinese product whose curriculum aligns to HSK 3. The learning agent reads progress, current exercise state and content context, then chooses the next practice, correction or prompt. Dialogue, practice and feedback become a controlled loop rather than a chat window with no memory.

We delivered the server API, business logic, data model and admin — learning paths, content, permissions and the controls the app runs on — and integrated with the partner client. The agent boundary is explicit: policy decides, content grounds, the runtime records state, and the human teaching team can change the system without a client release.

In the work

Stateful learning agent
Progress, exercise state and feedback give the agent enough context to choose the next move instead of repeating a generic prompt.
Grounded and operable
Content, permissions and learning data are managed behind an admin console, so teaching operations can adjust the agent without a release for every edit.
Agent behind a stable contract
The partner app is a client. The agent policy, state and business rules stay behind a server contract that can evolve independently.

The problem

A partner needed a from-zero HSK 3 Chinese product for international learners. The hard part was not adding chat; it was making an agent choose the next learning action from curriculum, learner state and feedback, while keeping a clean boundary to the partner-built app.

Measured result

Delivered the learning-agent substrate from 0 to 1: the product has a stateful next-action loop instead of a word-list stack, teaching operations can adjust content and inspect learning data, and the app remains decoupled from core agent logic.

Scope

Agent-facing server API, business logic and learner-state model
Curriculum, content, permissions and an operations console built from zero
Integration with the partner app through a stable client/server contract

What shipped

→ HSK 3 learning-agent loop: inspect state → choose practice → give feedback → advance
→ Interaction policy for pacing, correction, grounding and the next prompt
→ Content and learning-data control surface for the teaching team

Anonymisation

The partner, commercial data and contract name are not published. This page exposes only our technical scope and verifiable delivery outputs.

Work

Anonymised delivery

Agent for P0 scheduling and data sync

Anonymised delivery

DataClaw BI analysis agent

Anonymised delivery

Data-platform Agent runtime and MCP tool plane

Anonymised delivery

Governed knowledge gate for analysis Agents

Anonymised delivery

Alert-to-remediation operations Agent

Anonymised delivery

Agent control plane with guarded actions

Start a conversation