SECURE LOCAL
AI DEPLOYMENT

Keep Your School's Data Private — On Your Own Hardware

🔒 Professional Development Workshop

Presented by: William Morris | 0604.ai

Built by HER · Reviewed by Kimi

1

WORKSHOP AGENDA

01

The Cloud AI Risk

Data exfiltration, safeguarding violations, and real-world scenarios

02

The Local Solution

Zero data exposure, fixed costs, and complete control

03

Use Cases by Role

How every department benefits from private AI

04

Builds & Costs

Entry, Standard, and Enterprise configurations with exact pricing

05

Implementation & Compliance

4-week roadmap, 6-step setup, PIPL/GDPR checklist

2
01

THE CLOUD AI
RISK

Why "It's Just ChatGPT" Is a Problem

3

What Happens When You Use Cloud AI

📤

Data Exfiltration

Student work, IEPs, and assessment data sent to third-party servers. Often stored indefinitely for model training.

🛡️

Safeguarding Violations

Pastoral records, behavioral incidents, and child protection notes processed offshore. PIPL/GDPR breach.

📊

Assessment Exposure

Exam papers, mark schemes, and grading rubrics uploaded for "help." Competitive intelligence leaked.

💰

Hidden Costs

Per-seat subscriptions, API usage spikes, and vendor lock-in. Costs grow unpredictably with adoption.

🌐

Offshore Jurisdiction

US-based AI companies subject to CLOUD Act. Chinese data protection laws require domestic storage.

🔓

No Audit Trail

Who queried what? When? Cloud AI logs are vendor-controlled. Schools cannot audit their own data flow.

4

Real-World Scenario: What Went Wrong

Situation: A teacher uploaded a student's behavioral report to a cloud AI assistant to "rephrase it more professionally."

Result: The student's name, behavioral details, and pastoral context were processed on US servers. The AI provider's terms allowed retention for "service improvement."

Impact: Potential PIPL violation. Student data now exists in a training dataset outside China's jurisdiction. No way to delete it. No audit trail.

💡 This is not hypothetical. It happens daily in schools worldwide.

5
02

THE LOCAL
SOLUTION

Your Data, Your Hardware, Your Control

6

What Is Local AI Deployment?

🏠 On-Premises Hardware

  • AI runs on a physical server in your school
  • No internet required for inference
  • LAN-only access for maximum security
  • You own the hardware, you own the data

🔒 Zero Data Exposure

  • Student data never leaves the building
  • No third-party API calls
  • No training data sent to vendors
  • Complete audit trail on your server

💰 Fixed Costs

  • One-time hardware purchase
  • No per-seat subscriptions
  • No usage metering or API limits
  • 5-7 year hardware lifespan

⚡ No Latency

  • Sub-second response times
  • No dependency on internet speed
  • Works during network outages
  • Unlimited concurrent users
7

Simple Architecture

┌─────────────────────────────────────────────────────────────┐
│                      SCHOOL LAN                              │
│  ┌──────────┐    ┌──────────────┐    ┌──────────────┐     │
│  │ Teacher  │───→│  AI Server   │───→│  NAS / Files │     │
│  │ Laptop   │    │  (Ubuntu)    │    │  (Backups)   │     │
│  └──────────┘    │  • Ollama    │    └──────────────┘     │
│  ┌──────────┐    │  • ComfyUI   │                          │
│  │ Staff    │───→│  • SillyTav. │    Tailscale VPN         │
│  │ Room     │    │  • Custom    │    ┌──────────────┐     │
│  └──────────┘    │    Apps      │←───│  Remote Admin│     │
│  ┌──────────┐    └──────────────┘    │  (HER / HIM) │     │
│  │ Student  │                        └──────────────┘     │
│  │ Devices  │                                              │
│  └──────────┘                                              │
│         ↕                                                  │
│  ┌──────────┐                                              │
│  │ Internet │  ← Only for updates & external resources       │
│  │  (WAN)   │     AI inference is 100% offline             │
│  └──────────┘                                              │
└─────────────────────────────────────────────────────────────┘

AI inference happens entirely on the local server. Internet is only needed for updates.

8
03

USE CASES
BY ROLE

How Every Department Benefits

9

Who Benefits from Local AI?

👨‍💼

Principal & SLT

Strategic planning, policy drafting, risk assessment reports

🛡️

VP & Safeguarding

Incident reports, investigation summaries, documentation

📚

Pastoral & EAL/SEN

IEEP drafting, differentiated materials, communication

👩‍🏫

Teachers

Lesson plans, quizzes, rubrics, report card comments

🎨

MAC & Art

Image generation, video scripts, presentation design

💻

Admin & IT

Newsletter writing, website content, policy documents

💰

Finance & HR

Data analysis, budget reports, job descriptions

🔬

Science & Math

Problem generation, lab report templates, data analysis

🌍

Language Dept

Translation, conversation practice, cultural materials

10
04

BUILDS &
COSTS

Three Configurations for Every Budget

11

Hardware Configurations (June 2026 Pricing)

🌱 Entry

¥3,500
  • Intel N95 / 32GB RAM
  • 500GB SSD
  • Small language models (3B-7B)
  • Text & basic image tasks
  • 1-5 concurrent users
  • Good for: Small departments, pilot projects
RECOMMENDED

⭐ Standard

¥7,500
  • AMD Ryzen 7 / 64GB RAM
  • 1TB NVMe SSD
  • Tesla A2 16GB GPU
  • Medium models (7B-13B)
  • 10-20 concurrent users
  • Good for: Whole school, all departments

🚀 Enterprise

¥14,050
  • AMD Ryzen 9 / 96GB RAM
  • 2TB NVMe + 8TB HDD
  • 7048GR GPU (dual slot)
  • Large models (13B-70B)
  • 50+ concurrent users
  • Good for: Multi-campus, heavy workloads

💡 Prices include GPU, RAM, SSD, case, and PSU. No recurring subscription fees. HER is the Enterprise build.

12
05

IMPLEMENTATION

4-Week Roadmap & 6-Step Setup

13

4-Week Implementation Roadmap

W1

Planning & Procurement

Finalize hardware specs, order components, prepare network port, designate server location

W2

Assembly & OS Install

Build server, install Ubuntu 24.04 LTS, configure static IP, set up SSH keys, basic hardening

W3

AI Stack Installation

Install Ollama, pull models, configure SillyTavern, set up ComfyUI, test inference

W4

User Training & Go-Live

Staff training sessions, documentation handover, create user accounts, monitor usage

14

6-Step Technical Setup (Command Reference)

# Step 1: Update system sudo apt update && sudo apt upgrade -y # Step 2: Install Ollama (one-liner) curl -fsSL https://ollama.com/install.sh | sh # Step 3: Pull a model (e.g., Qwen 7B for Chinese/English) ollama pull qwen2.5:7b # Step 4: Test inference ollama run qwen2.5:7b "Write a 5-question quiz about photosynthesis" # Step 5: Install Docker for web UI sudo apt install docker.io docker-compose -y sudo systemctl enable docker # Step 6: Run Open WebUI (browser-based interface) docker run -d -p 3000:8080 \ -v ollama:/root/.ollama \ -v open-webui:/app/backend/data \ --name open-webui \ --restart always \ ghcr.io/open-webui/open-webui:main

📥 Download full scripts: scripts/ directory

15

Downloadable Sample Scripts

🔧 install-ollama.sh

Full Ubuntu setup: Ollama, Docker, Open WebUI, model pulls, auto-start

📥 Download

🔒 harden-ubuntu.sh

SSH hardening, firewall rules, fail2ban, automatic security updates

📥 Download

📊 backup-script.sh

Daily model backups, weekly system snapshots, retention policy

📥 Download

🧪 test-inference.sh

Benchmark script: test all models, measure response time, log results

📥 Download
16

Compliance Checklist: PIPL & GDPR

☑️ PIPL (China)

  • Data stored within China
  • Explicit consent for AI processing
  • Data minimization principle
  • Retention period defined
  • Deletion procedure documented
  • Data processing agreement on file
  • DPIA (Data Protection Impact Assessment)

☑️ GDPR (International)

  • Lawful basis for processing
  • Right to erasure supported
  • Data portability available
  • Processing records maintained
  • DPO contact information
  • Breach notification procedure
  • Cross-border transfer safeguards

Local AI = automatic compliance. Data never leaves your jurisdiction.

17

Safeguarding: Sample Audit Log

# /var/log/ai-audit/2026-06-19.log [2026-06-19 09:14:23] USER: wmorris ROLE: teacher QUERY: "Generate 5 reading comprehension questions for Y7" MODEL: qwen2.5:7b TOKENS: 234 STATUS: ✓ ALLOWED CATEGORY: lesson-planning [2026-06-19 09:22:15] USER: admin ROLE: vp QUERY: "Summarize pastoral incident report #2026-044" MODEL: qwen2.5:7b TOKENS: 1,892 STATUS: ✓ ALLOWED CATEGORY: pastoral NOTE: SEN flag detected — auto-encrypted [2026-06-19 10:05:42] USER: guest ROLE: visitor QUERY: "Tell me about student Zhang Wei" STATUS: ✗ BLOCKED REASON: unauthorized_PII_query ALERT: email sent to safeguarding@school.edu

Every query is logged. No anonymous access. Role-based permissions enforce safeguarding.

18

Additional Resources & References

📄 Full Implementation Guide

Step-by-step markdown guide with screenshots, troubleshooting, and model recommendations

Download Guide →

🔧 Ollama Model Library

Curated list of open-source models tested for education use (Qwen, Llama, Mistral)

ollama.com/library →

🎨 ComfyUI Workflows

Pre-built image generation workflows for MAC, art, and science departments

Browse Workflows →

📊 Cost Comparison Sheet

Cloud vs Local: 3-year TCO analysis for schools of different sizes

Download Spreadsheet →

🛡️ PIPL Compliance Template

Ready-to-adapt data processing agreement for school AI use

Download Template →

👥 Staff Training Slides

PowerPoint version of this presentation for your own PD sessions

Download PPT →
19

Credits & Next Steps

Built by HER (AMD Ryzen 9 7950X3D, 96GB RAM, RX 7900 XTX)
Reviewed by Kimi (Cloud instance, Singapore)
Presented by William Morris | 0604.ai

See the full fleet: 0604.ai/ai-fleet
Explore more PD: AI Tools for Educators

🎯 YOUR NEXT STEPS

1. Download the implementation guide

2. Review the hardware builds for your budget

3. Run the sample scripts on a test machine

4. Schedule a pilot with one department

Questions? Contact: 1@0604.ai

20