AI
AI Maths Tutoring Platform
A live AI tutoring room where a self-hosted language model guides students with hints, and parents or teachers can observe remotely.
Overview
An education company had a maths tutor model trained on their syllabus and needed a real product around it: a tutoring room students actually use, deployed on infrastructure they control.
Client challenge
Small models tend to blurt out answers, CPU hosting made every reply take 40 seconds or more, and parents wanted to watch sessions from another device.
Solution
We built the tutoring room in Next.js and engineered the prompts so the model asks one guiding question instead of solving the problem. We moved inference to an on-demand AWS GPU server and added a remote observer mode over WebRTC.
Features
What the product includes
Tutor, student and observer roles
Hint-first tutoring that avoids giving away answers
Remote observer joins with a 6-character code
Voice typing for students
Safeguarding: sensitive messages skip the model and alert the observer
Diagnostic test with lead capture and admin export
Process
How this project was delivered
- 1
Deployed the model with Docker on AWS EC2
- 2
Built the room against a mock tutor so UI work never waited on the server
- 3
Tuned prompts against real model behaviour
- 4
Load-tested on the live GPU deployment
Results
Outcomes that mattered
- Reply time cut from about 40 seconds on CPU to 1 to 2.6 seconds on GPU
- Remote observing works across devices and networks
- Hosting cost controlled by running the GPU only when needed
Have a project like this in mind?
We will help you clarify scope, stack, and timeline before development begins.