The Challenge: 85,000 Students and No Tutor Available at 11 PM
A UK-based online learning platform provides GCSE maths revision to 85,000 students. Usage data reveals that 62% of study sessions happen between 7 PM and midnight — well outside the hours when live tutors are available. Students hit a concept they do not understand, click through three static explanation pages, remain confused, and abandon the session. The average session dropout rate after encountering a difficult topic is 45%. The platform’s educational team knows that personalised, conversational explanation — adapting language and examples to each student’s level — is the proven approach, but hiring enough tutors to cover 85,000 students during peak evening hours is financially impossible.
Third-party AI tutoring APIs pose significant concerns. Student interaction data — including learning difficulties, assessment scores, and personal details — constitutes children’s data under UK GDPR and the Children’s Code. Sending this data to external API providers, particularly those based outside the UK, creates regulatory risk the platform cannot accept. Data must remain on UK soil.
AI Solution: Self-Hosted LLM Tutor
A self-hosted open-source LLM fine-tuned on maths pedagogy provides conversational tutoring through the platform’s chat interface. The model is trained on 50,000 exemplar tutor-student conversations covering the GCSE maths syllabus, learning to explain concepts using analogies, step-by-step worked examples, and Socratic questioning rather than simply providing answers. It adapts its language complexity based on the student’s year group and demonstrated understanding.
The tutor runs on a dedicated GPU server with vLLM, handling hundreds of concurrent student conversations with sub-second response times. A retrieval-augmented generation layer connects the model to the platform’s syllabus content, ensuring explanations align with the specific exam board specification the student is studying.
GPU Requirements
Peak evening usage sees 3,000-5,000 concurrent student sessions. Each conversation requires low-latency generation — students expect tutor responses within 2 seconds. A 7B model fine-tuned for maths tutoring provides the quality-speed balance needed for this workload.
| GPU Model | VRAM | Response Latency (7B model) | Concurrent Students |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~0.9 seconds | ~800 |
| NVIDIA RTX 6000 Pro | 48 GB | ~1.1 seconds | ~1,200 |
| NVIDIA RTX 6000 Pro | 48 GB | ~0.8 seconds | ~1,500 |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~0.5 seconds | ~2,500 |
For 5,000 concurrent students at peak, a pair of RTX 6000 Pros or three RTX 6000 Pro GPUs handles the load. For initial rollout with lower concurrency, a single RTX 6000 Pro provides excellent performance. All options through GigaGPU’s private AI hosting guarantee UK data residency for children’s data compliance.
Recommended Stack
- vLLM for high-throughput LLM serving with continuous batching and streaming output.
- Mistral 7B or LLaMA 3 8B fine-tuned on pedagogical conversation data using LoRA.
- LlamaIndex for retrieval-augmented generation connecting the tutor to syllabus-specific content.
- Guardrails layer ensuring the model stays on-topic, refuses to complete homework directly, and maintains age-appropriate language.
- WebSocket API for streaming tutor responses to the platform’s chat widget.
For handling handwritten maths from student uploads — photos of working on paper — add a vision model to parse handwritten equations and identify where the student made an error. Integrate document AI to process past paper PDFs into the retrieval system.
Cost Analysis
Third-party AI tutoring APIs charge £0.02–£0.08 per interaction. With 85,000 students averaging 15 interactions per session and 3 sessions per week, monthly API costs would reach £153,000–£612,000. Self-hosting on dedicated GPUs provides unlimited interactions at a fixed monthly cost that represents a tiny fraction of those figures. The platform can offer unlimited AI tutoring without worrying about per-interaction economics.
The educational impact justifies the investment independently. Early pilot data shows a 23% improvement in topic mastery rates among students using the AI tutor compared to those relying solely on static content. For a platform competing on learning outcomes, this is the differentiator that drives subscriber retention.
Getting Started
Assemble your training dataset: compile 10,000 high-quality tutor-student conversations from your existing live tutoring sessions (anonymised). Fine-tune a 7B model using LoRA, testing output quality against a rubric scored by your educational team. Deploy to a beta group of 1,000 students for two weeks, measuring engagement metrics and learning outcome improvements before full rollout.
GigaGPU provides UK-based dedicated GPU servers with guaranteed data residency for education workloads. Add an AI chatbot for administrative support queries, or scale GPU allocation during exam season peaks.
GigaGPU offers dedicated GPU servers in UK data centres with full GDPR compliance for children’s data. Launch personalised AI tutoring on private infrastructure today.
View Dedicated GPU Plans
