How a Shenzhen School Deployed AI Teaching Assistants—While Keeping 2,000 Students' Data 100% Local
In spring 2026, the IT department of a Shenzhen key middle school did something unusual: they brought AI into the computer lab—not via the cloud, but locally on two KAIHE boxes.
Why "unusual"? Because China's education sector has a conflicted relationship with AI. Teachers want AI-assisted grading and tutoring, but uploading student data to the cloud crosses a red line. China's Personal Information Protection Law (Article 28) explicitly requires parental consent for processing personal data of minors under 14. The essays, test answers, and learning patterns of 2,000 students—no school dares send that to OpenAI's servers.
The Setup: 2x KAIHE C1 + OpenClaw Teaching Agents
The school deployed two KAIHE C1 units on the internal network:
- C1 #1: Running Qwen2.5-7B INT4, handling the "Essay Grading Agent"
- C1 #2: Running DeepSeek-R1-7B INT4, handling the "Math & Science Tutoring Agent"
Connected via OpenClaw Workflow: student submits work → Essay Agent grades it → difficult questions auto-forwarded to Math Agent → personalized feedback report generated.
Results
Essay Grading Agent
- Speed: 800-word essay graded with feedback in ~30 seconds
- Quality: Cross-validated against two Chinese teachers—scoring deviation within ±3 points (100-point scale); feedback rated "better than most teaching assistants"
- Time saved: Weekly grading burden reduced from 120 essays to 40 reviews—67% time savings
Math & Science Tutoring Agent
- Accuracy: ~92% correct on middle school math (honestly says "I need more time" on out-of-scope problems)
- Usage: 120 daily queries after 2 months (mostly during breaks and after school)
- Most popular feature: Not giving answers—instead providing step-by-step hints. When a student is stuck, the Agent suggests just the next step's approach, never the final answer.
Privacy Protection
- 100% of data stays on the school's internal LAN, completely isolated from the internet
- Student names and IDs are auto-anonymized before Agent input (replaced with random IDs)
- Grading results and tutoring logs automatically backed up to local NAS daily—zero cloud dependency
Cost Comparison
| Approach | Monthly Cost | Data Security | Customization |
|---|---|---|---|
| GPT-4o API (pay-per-use) | ~$1,100/mo | ❌ Data leaves China | Low (prompt-level only) |
| Domestic LLM API | ~$400/mo | ⚠️ Data at third party | Low |
| KAIHE C1 × 2 (one-time) | $0/mo | ✅ Fully local | High (Agent + Workflow) |
Hardware cost: 2× C1 = ¥1,998 (one-time). Electricity is negligible (each unit draws <15W).
Bottom Line
When AI enters the classroom, "good enough" matters more than "most powerful." The C1's 7B model handles teaching agents perfectly, and the privacy assurance of local deployment is something no cloud solution can offer.
The school proved one thing: the first mile of AI in education isn't compute—it's trust.