Project Story
Modern educational platforms tend to measure whether a student gets an answer right or wrong, but they rarely understand why. We wanted to create something that adapts to the learner rather than expecting every learner to adapt to the platform.
Project PRISM was built around the idea that education should reflect how cognitive science shows people actually learn. Instead of acting as another AI question generator, we wanted to build the foundations of an adaptive learning system that identifies misconceptions, adjusts the presentation of content, and supports accessibility from the start.
Our proof of concept focuses on Sixth Form students, but the architecture is designed so it could eventually support learners of all ages and abilities.
⸻
Inspiration
Our inspiration came from several areas of research.
We explored cognitive psychology, memory science, adaptive learning systems, knowledge tracing, accessibility research and psychometrics. Research into the spacing effect, retrieval practice, cognitive load theory, metacognitive calibration, Item Response Theory, Knowledge Graphs and modern adaptive learning algorithms showed us that many existing educational platforms still rely heavily on static question banks and simple percentage scores.
We wanted to combine these ideas into a single platform that continually asks:
“What is the student struggling with?”
rather than
“Did they get this question right?”
This shift in thinking became the foundation of Project PRISM.
⸻
What it does
Project PRISM is an AI assisted adaptive learning platform that analyses much more than a student’s final answer.
Our proof of concept records correctness, confidence and concept information before using Gemini to provide personalised feedback and identify likely misconceptions.
Rather than moving through questions in a fixed order, the platform models concepts as connected knowledge. If a learner struggles with an advanced topic, the system can recommend revisiting prerequisite concepts instead of simply repeating similar questions.
Accessibility is also integrated directly into the interface. Rather than treating accessibility as an afterthought, the UI can adapt its typography, spacing and presentation style to better support different learners while maintaining a clean, distraction free experience.
Although the current prototype demonstrates these ideas on a smaller scale, the architecture is designed to expand into a fully adaptive learning platform.
⸻
How we built it
The frontend was built using modern web technologies with an emphasis on clarity, accessibility and responsive design.
The backend manages user interactions, question data and communication with the Gemini API.
We structured educational content around concepts rather than isolated questions, allowing the platform to reason about relationships between topics instead of treating every question independently.
Gemini is not simply used to mark answers. Instead, it acts as an educational reasoning engine, generating personalised feedback and suggesting appropriate next steps based on the student’s performance.
Throughout development we continually referred back to educational research to justify design decisions, ensuring that each feature had a cognitive or pedagogical rationale rather than existing simply because it was technically possible.
⸻
Challenges we ran into
One of the biggest challenges was balancing ambition with the limited timeframe of a hackathon.
Many of the cognitive models we researched, including Deep Knowledge Tracing, adaptive spaced repetition algorithms and psychometric models such as Item Response Theory, would normally require large datasets collected over weeks or months before producing meaningful results. As a result, we had to simplify these ideas into a proof of concept while preserving the overall architecture.
Another challenge was preventing AI from becoming the entire product. Rather than relying on Gemini to simply generate questions, we focused on using it as an educational reasoning engine that could explain misconceptions and personalise learning pathways.
We also spent considerable time simplifying complex educational research into an interface that remained intuitive and accessible without overwhelming users.
⸻
Accomplishments we’re proud of
We’re particularly proud that the project is driven by educational research rather than trends.
Instead of building another AI chatbot for education, we combined ideas from cognitive science, memory research, adaptive learning, accessibility and knowledge modelling into a coherent proof of concept.
From a technical perspective, we successfully integrated AI into a structured workflow rather than allowing it to make arbitrary decisions. The platform separates educational content, learner state and AI reasoning, making it scalable and maintainable.
We’re also proud that accessibility was considered throughout the design process instead of being added afterwards. Features such as adaptable layouts, clearer typography and personalised presentation demonstrate how educational software can be more inclusive without compromising usability.
Finally, despite the limited development time, we’ve built a strong foundation for a system that could eventually support adaptive revision, personalised feedback, intelligent prerequisite detection, knowledge graphs, confidence calibration and scientifically informed learning pathways. Our goal was not simply to build another revision website, but to demonstrate a different way educational software could work.
