Three Ways to Deploy Classroom AI on Interactive Flat Panels: On-Premise, Cloud, or Hybrid
Riotouch Cloud deployment models, explained for school IT teams, integrators and procurement
Quick answer. Riotouch runs the same platform in three deployment modes. Mode 1 — On-Premise Cloud, where all compute and data stay inside the school's own network. Mode 2 — Cloud (Networked) mode, hosted in an EU region. Mode 3 — Hybrid, where student data never leaves the campus and only redacted workloads travel to the cloud. Which one you use comes down to three questions: how reliable is the campus network, what do the data-protection rules require, and how many users does the deployment serve? Sizing starts at a single rack server for a sub-200-student school and scales to a multi-campus fleet running 400B-class models.
Riotouch is an interactive flat panel manufacturer with 16+ years in the category and deployments in 80+ countries — which is why the question we get asked most often is not "what does the panel do?" but "where does the software live, and where does our students' data go?"
1. What "the cloud" actually means on an IFP deployment
Most panel vendors sell you a dashboard. What matters underneath is a stack. Riotouch's platform is seven layers deep, and the deployment mode only changes where layers 1–3 run and where AI inference happens. Everything above them behaves identically in all three modes.
| Layer | What it does | Changes with deployment mode? |
|---|---|---|
| 7 — Client | Android app for students/parents, Web PWA for teachers and admins, optional iOS app, smart board terminal (Android/Windows), recording & meeting hardware | No |
| 6 — Application | AI generation, exam solving, 3D modelling, interactive 3D courseware, writing correction, speaking practice, lab simulation, coding tutor, class observation, textbook assistants, graded reading, translation, adaptive learning, question generation, learning companion, AI search | No |
| 5 — Cloud services | DMS device management, CMS content management, RMS resource management, LMS teaching management, live streaming, AI-Gen platform, home-school portal, whiteboard, recording, meeting, exam, notification and analytics services | Runs on-premise, in the cloud, or split |
| 4 — AI engine | Local AI Engine (data stays on campus) and Cloud AI Engine (redacted data only), behind an AI Gateway | Yes — this is the real choice |
| 3 — Data | PostgreSQL 16 + pgvector, Redis 7, MinIO (S3) object storage, Meilisearch full-text search | Same software, different location |
| 2 — Runtime | Docker, Kubernetes, Nginx, Prometheus + Grafana, Loki + Promtail, GitLab CI + ArgoCD | No |
| 1 — Operating system | Ubuntu 22.04 LTS / Debian 12, NVIDIA 550+ / Intel / AMD drivers, CUDA Toolkit 12.x with cuDNN and TensorRT | No |
Two things follow from that table. First, a school never trades features for sovereignty — the parent app, the teacher whiteboard, the device-management dashboard and the district analytics all work the same on-premise as in the cloud. Second, moving between modes is a hosting change, not a re-implementation.
2. Mode 1 — On-Premise Cloud (local deployment)
Everything runs on a server inside the school. Content, accounts, recordings, exam data and model weights never leave the building. In the platform's own words: Local AI Engine — data stays on campus.
Sizing ladder: Plan A1, A2, A3
| Plan A1 — Light | Plan A2 — Standard | Plan A3 — Flagship | |
|---|---|---|---|
| Scale | <200 students | 200–1,000 students | 1,000+ students / multi-campus |
| CPU | Intel i9-14900K or AMD Ryzen 9 7950X (24C/32T) | AMD Threadripper PRO 7965WX (24C) or Xeon w5-3435X | 2× AMD EPYC 9454 (48C ×2 = 96C) or 2× Xeon Gold 6448Y |
| Memory | DDR5 64GB | DDR5 ECC 128GB | DDR5 ECC 256–512GB |
| GPU | 1× RTX 4090 24GB | 2× RTX 4090, or 1× L40S 48GB | 4× A6000 48GB, or 2× H100 80GB |
| Storage | 2TB NVMe Gen4 + 8TB NAS (RAID5) | 4TB NVMe Gen4 + 24TB NAS (RAID6) | 8TB NVMe RAID10 + 100TB+ Ceph |
| Network / power | 2.5GbE, 1KVA UPS | 10GbE, 3KVA online UPS | 25GbE + SD-WAN, 6KVA UPS + generator |
| Local model capacity | 14B Q4 | 32B Q4 / 70B Q4 | 72B FP16 / 405B Q4, multi-model |
Classroom-level option: Edge AI Box
Where a school wants inference sitting next to the panel rather than only in the server room, an Edge AI Box (Jetson Orin Nano/NX, one per N classrooms) handles low-power local inference with PoE power — useful for handwriting recognition, board-face tracking and small-model responses with no server round-trip.
Room requirements (integrator note)
Plan for: precision HVAC, anti-static flooring, physical access control, smoke and water-leak sensors, KVM, and a 42U rack. These are the items that turn "a server in a closet" into a supported deployment — and they are the ones most often left out of the BOM.
What runs locally in this mode
DMS device management (registration, monitoring, OTA, alerting, asset audit), CMS publishing and review with AI tagging and translation, RMS resource management with semantic search, LMS teaching management with SCORM / LTI 1.3, whiteboard with handwriting recognition, recording with AI editing and chapter marking, WebRTC video meetings, online exams with anti-cheat and AI proctoring, notifications (email/SMS/push/webhook), and analytics dashboards.
What the Local AI Engine can do
LLM inference via Ollama/vLLM with Qwen2.5-class models; RAG with LangChain and pgvector search; agent frameworks (LangGraph, ReAct / tool calling); speech with Whisper ASR and VITS2/Piper TTS; OCR and computer vision (PaddleOCR, board-face recognition); bilingual CN/EN embeddings; optional image generation (SD/SDXL); and a model registry with canary release and rollback.
3. Mode 2 — Cloud (Networked) mode
The same platform, hosted. For EU-region deployments the reference layout is:
| Infrastructure | Services | Best for | |
|---|---|---|---|
| B1 | AWS Frankfurt (eu-central-1) | EC2 / RDS / S3 / CloudFront | Elastic compute and CDN |
| B2 | Azure EU West (Netherlands / Ireland) | AKS / CosmosDB / Azure OpenAI | AI inference, hybrid AD integration |
| B3 | OVHcloud / Hetzner / Scaleway | EU-native, GDPR-friendly hosting | Low-cost storage and backup |
| B4 | Mistral AI / Aleph Alpha | European AI companies — data stays in the EU | LLM inference and compliance |
Cloud AI Engine — redacted data only
Azure OpenAI (EU region), Mistral AI (France) and AWS Bedrock (EU region) cover advanced reasoning, creative generation and multimodal work. SOC 2 and GDPR agreements are signed, the EU Data Boundary is enforced, and usage is metered per request.
What the cloud is actually used for
Five sanctioned use cases — and they all run on redacted data only:
- Heavy reasoning when the local model is not sufficient
- Elastic scaling for exam-season and enrolment peaks
- Static assets and video delivery
- Encrypted backups
- SFT / LoRA model training
"What happens if the internet drops?" On a cloud-only deployment, management and new content push pause; playback continues from local cache. On the on-premise and hybrid configurations, the platform keeps running. That difference is the reason hybrid exists.
4. Mode 3 — Hybrid (local + networked)
Hybrid is not "half on-premise". It is a routing policy executed by the AI Gateway, whose five jobs are: smart routing, PII redaction, audit logging, degradation, and token metering. Every request is classified before anything leaves the school.
| Workload | Stays on campus | May go to the cloud |
|---|---|---|
| Student names, faces, voices, handwriting, exam answers, grades, live board content | Always | Never — only redacted derivatives |
| Routine AI tutoring, OCR, TTS, RAG over school-owned content | Yes, on the local model | — |
| Heavy reasoning / large-model tasks when local capacity is insufficient | — | Yes, redacted only |
| Elastic peaks (exam season, enrolment drives) | Baseline capacity | Burst capacity |
| Video and static asset delivery | Local cache | CDN |
| Backup and disaster recovery | 3-2-1 local copy | Encrypted offsite copy |
| SFT / LoRA fine-tuning | Optional | Redacted data only |
Integrator note. In a hybrid deployment, ask two questions before signing off: who owns the redaction policy, and is the audit log exportable? The AI Gateway logs every routed request, so a school can show exactly which data left the campus, when, and under which rule.
5. Choosing your mode
| Situation | Recommended mode |
|---|---|
| Single campus, under 200 students, decent LAN | Mode 1 — Plan A1 |
| 200–1,000 students, several buildings, IT staff on site | Mode 1 — Plan A2, or Mode 3 with an A2 baseline |
| Multi-campus / district, 1,000+ users, one dashboard across sites | Mode 3 — Plan A3 + SD-WAN |
| Data-protection rules prohibit student data leaving campus | Mode 1, or Mode 3 with redacted-only cloud usage |
| Reliable internet, no server room, small IT team | Mode 2 |
| Sharp load spikes during exam seasons | Mode 3 — local baseline plus cloud burst |
| No on-site IT at all, maintenance must be zero-effort | Mode 2 |
6. Security and data protection — identical in all three modes
- Data classification: Public | Internal | Confidential | Highly Confidential (PII)
- Backup policy: 3-2-1, encrypted storage, auto-expiry, offsite disaster recovery
- Network and security: enterprise UTM firewall and WAF, zero trust, SD-WAN for multi-campus, Wi-Fi 6E/7, VLAN isolation, full-link TLS 1.3, IDS/IPS
- OS hardening: AppArmor/SELinux mandatory access control and sandboxing, unattended security upgrades with auto-patch and reboot windows
- Agent roles covered by the platform: student, teacher, parent, admin, and counsellor (local model with human fallback)
7. Key takeaways
- Riotouch ships one platform in three deployment modes — on-premise, cloud, and hybrid. Apps, portals and management features are identical; only the hosting location and the AI routing policy change.
- On-premise scale is a three-step ladder: A1 for under 200 students running 14B-class models, A2 for 200–1,000 students at 32B–70B, and A3 for multi-campus fleets at 72B FP16 or 405B Q4 with dual GPUs and Ceph storage.
- In cloud mode, the EU-region layout pairs AWS Frankfurt for compute and CDN, Azure EU West for AI inference and AD integration, EU-native hosts for storage and backup, and European AI vendors (Mistral, Aleph Alpha) for compliant inference.
- Hybrid's rule is one sentence long: student data never leaves the campus; only redacted workloads go to the cloud, and every routed request is logged.
- Going hybrid is not a licence to send everything upward. There are five sanctioned cloud use cases: heavy reasoning, elastic peaks, static assets, encrypted backups and SFT/LoRA training.
- Room infrastructure — HVAC, anti-static flooring, access control, sensors, KVM and a 42U rack — is part of the deployment, not an afterthought. Put it in the BOM.
Planning a classroom AI deployment?
Tell us your campus size, student count and data-residency rules — we will map them to on-premise, cloud or hybrid and give you the server plan.
See Riotouch CloudFrequently asked questions
What does classroom AI deployment mean on an interactive flat panel?
It means the teaching software, device management, content publishing and AI models that run on the panel are hosted somewhere: on a server inside the school (on-premise), in a hosted environment (cloud), or split between the two by a routing policy (hybrid). The panel itself is the same in all three cases.
Can student data stay on campus?
Yes. In the on-premise mode every record — accounts, content, recordings, exam data and model inference — stays on the school server. In hybrid mode the routing rule is one sentence long: student data never leaves the campus, and only redacted workloads are ever sent upward.
What happens if the internet drops?
On a cloud-only deployment, management and new content publishing stop and teaching falls back to local apps already installed on the panels. On-premise and hybrid deployments keep management, publishing and local AI inference running, because the core services sit on the school server.
How big a server does a school need for on-premise AI?
Riotouch uses a three-step ladder: Plan A1 for under 200 students running 14B-class models, Plan A2 as the standard configuration for a typical campus, and Plan A3 as the flagship configuration for large or multi-building deployments. Where inference is needed next to the panel rather than in the server room, an Edge AI Box is added at classroom level.
Is the security model different between the three modes?
No. Data classification (Public / Internal / Confidential / Highly Confidential), the 3-2-1 encrypted backup policy, network protection, OS hardening and the agent roles covered by the platform are identical in on-premise, cloud and hybrid deployments. What changes is where the workloads run, not how they are protected.
Related Product
Riotouch Cloud Solution — one platform, three deployment modes for interactive flat panels. Hardware: RK3588 Smart Board · OPS modules · Edge AI Box.