tensor paper

Inspiration

TensorPaper began with something I had seen personally.

When I left secondary school in Scotland six years ago to study Artificial Intelligence at university, classrooms already contained laptops, tablets and interactive displays. However, much of that technology was being used as a digital replacement for paper rather than to give students genuinely personalised support.

I also saw pupils being left behind, particularly those with learning disabilities or additional support needs. This was not because their teachers did not care. Teachers simply did not have enough time to adapt every explanation, create individual practice material, mark every response quickly and identify exactly why each pupil was struggling.

At the end of university, I completed a Computing Science teaching course and returned to the classroom from a different perspective. The problem had not improved. If anything, the pressure on teachers had become more visible.

Scotland has experienced a serious decline in specialist Computing Science provision. The number of full time equivalent Computing Science teachers fell by 28% between 2008 and 2024/25. A separate 2024 investigation found that 66 Scottish secondary schools had no qualified Computing Science teacher, affecting more than 32,000 pupils, approximately one in eight secondary pupils.

This is part of a wider challenge. UNESCO estimates that the world will require another 44 million primary and secondary teachers by 2030. Technology is already present in many classrooms, but research shows that its educational impact remains uneven and depends heavily on teacher preparation, appropriate design and clear learning objectives.

I wanted to explore how AI could support teachers without attempting to replace them, while also giving students meaningful feedback when a teacher cannot be beside them.

That became TensorPaper.

What TensorPaper does

TensorPaper is a study support platform that allows students to complete real past paper questions digitally and receive immediate, explainable feedback.

The current version focuses on Scottish National 5, Higher and Advanced Higher Computing Science, but the underlying system is designed to expand to other subjects and examination boards.

Rather than presenting AI as an unquestionable teacher, TensorPaper creates a simulated peer marking experience. The student receives a suggested mark, sees which marking points were met or missed and is encouraged to compare the feedback critically with the official marking instructions.

This is important because the student is not just being told whether an answer is correct. They are encouraged to think like a marker, understand the success criteria and consider how their answer could be improved.

Research into genuine peer assessment suggests that reviewing work against assessment criteria can improve students’ understanding and academic performance. TensorPaper uses this principle as inspiration, while making it clear that the feedback comes from AI and may still require human judgement.

GPT 5.6 considers the question, the official marking instructions, any additional guidance and the student’s response. It then produces:

• An estimated mark
• The marking points the student achieved
• An explanation of anything that was missing
• Constructive feedback on how to improve

Turning Answers into Useful Educational Data

The longer-term value of TensorPaper goes beyond automatic marking.

Because pupils enter their answers digitally, every response can be connected to its question, qualification level, curriculum topic and individual marking points.

At present, a large amount of valuable evidence is written in exercise books, marked once and then rarely analysed again. TensorPaper can turn those responses into structured learning data at several levels.

Student Level

A pupil could see:

• Their strongest and weakest topics
• Misconceptions that repeatedly cost them marks
• The types of questions they find most difficult
• How their answers are improving over time
• Which areas they should revise next

Teacher and Class Level

A teacher could identify:

• Concepts that a large proportion of the class has misunderstood
• Individual pupils who may need additional support
• Marking points that are repeatedly being missed
• Whether a problem comes from missing knowledge or difficulty explaining that knowledge
• Which areas may need to be retaught using a different approach

School Level

A school could examine:

• Differences between classes and year groups
• Topics where additional resources may be required
• The effects of new teaching materials or interventions
• Areas where non-specialist teachers may need support
• Whether pupils are receiving consistent opportunities across the school

Exam Board and National Level

With appropriate consent, access controls, data minimisation and anonymisation, aggregated responses could help an exam board or education organisation understand:

• Which curriculum concepts pupils struggle with most often
• Why pupils lose marks on particular questions
• Whether marking guidance is being understood consistently
• Where teaching resources or professional learning may need to improve
• How experiences differ between schools, regions and socioeconomic groups

The purpose is not to collect more data for its own sake. It is to turn pupils’ answers into practical evidence that can improve how they are supported.

How I built it

TensorPaper has a React and TypeScript student interface for navigating real past paper questions, submitting answers and reviewing feedback.

One of the largest challenges was preparing the source material.

Question papers and marking instructions are designed for people, not software. A single question may continue over several pages and include diagrams, tables, code, introductory context and additional marking guidance stored elsewhere in the document.

For TensorPaper to mark reliably, each question must be converted into a structured format containing:

• The question and each subquestion
• Its page location
• Relevant diagrams and code
• The maximum available mark
• Individual marking points
• Additional marking guidance
• Context shared between multiple parts
• Its course topic and subsection

Preparing this manually took roughly three to five hours for each paper. For the initial collection of 15 papers, that would have represented approximately 45 to 75 hours of manual preparation.

I used Codex and GPT 5.6 to build, test and improve an extraction pipeline that converts the original question papers and marking instructions into structured data. Codex helped diagnose extraction failures, improve the schema, write validation tools and repeatedly test the process against difficult documents.

The completed collection of 15 papers required approximately 15 minutes of final human verification. This reduction is what makes it realistic to expand TensorPaper beyond one qualification or examination board.

Expanding to Other Subjects and Exam Boards

The Codex-assisted extraction system is one of the most important parts of TensorPaper because it makes expansion practical.

TensorPaper is not hard-coded around the exact layout of one Higher Computing Science paper. The extraction pipeline converts each paper into a reusable, structured format that describes:

• The question
• Supporting context
• Diagrams
• Marking points
• Curriculum topics

To add a new subject or exam board, the system can be given its question papers and marking instructions. Codex can then help adapt the extraction and validation rules to the new document structure, while GPT-5.6 converts the content into the same standardised question format used by TensorPaper.

A human still verifies the final output, but they no longer need to manually copy and organise every question, diagram and marking instruction.

This turns expansion into a repeatable process:

1. Add the new question papers and marking instructions.
2. Use the Codex-assisted pipeline to extract and structure them.
3. Automatically identify missing or inconsistent content.
4. Perform targeted human verification.
5. Add the verified questions to TensorPaper.

Without this extraction system, expanding across several subjects and exam boards could require hundreds or thousands of hours of manual preparation. With Codex, the same underlying process can be reused and improved as more document formats are encountered.

This is what makes it realistic for TensorPaper to begin with Scottish Computing Science and later support subjects such as Mathematics, Biology, Physics, Business and Geography, as well as qualifications from other UK and international exam boards.

Benchmarking with Codex

I did not want to judge the marker using a few carefully selected examples, so I began developing a larger benchmarking system.

Codex helped me create the evaluation pipeline, generate and organise test cases, send requests at scale, compare the returned marks with reference results and surface disagreements for inspection.

It also made the process considerably less stressful.

Long benchmark runs can fail because of API rate limits, temporary errors, interrupted connections or malformed model outputs. Codex helped build the queuing, concurrency controls, retry logic, validation, progress saving and recovery systems needed to handle these problems automatically.

This meant I did not have to constantly watch the benchmark or manually restart it whenever something went wrong. It could respect the API limits, continue processing safely and scale from a few answers to much larger evaluation sets.

I only reached the beginning of this benchmark work in the submission video, so I was unable to explain the full evaluation system, the adaptive reasoning approach or the planned data analytics in detail.

Adaptive reasoning

Not every exam question requires the same amount of reasoning.

A simple one mark recall question should be marked quickly, while an extended explanation, algorithm or programming question may require deeper analysis.

TensorPaper is being designed to adapt the reasoning effort to the complexity of the question. This should reduce unnecessary cost and waiting time for simple responses while allowing more difficult answers to receive the attention they need.

The system is also intended to recognise uncertain cases and recommend human review instead of presenting every AI judgement with false confidence.

Challenges I faced

The largest technical challenge was not building a text box connected to a model. It was making the entire process structured, repeatable and verifiable.

Questions can cross pages, diagrams may be essential, marking instructions contain exceptions and two student answers can express the same understanding in completely different language.

Another challenge was balancing automation with trust. In education, a convincing explanation is not enough if the awarded mark is wrong. I therefore focused on showing the marking instructions, connecting feedback to individual marking points and allowing uncertain cases to be reviewed.

The analytics also introduce important privacy questions. Any future class, school or national analysis would need clear access controls, appropriate data minimisation and anonymisation. The purpose should always be to support students and teachers, not to create another opaque system for ranking them.

What I learned

The most important lesson was that useful educational AI depends as much on data preparation and evaluation as it does on the model.

A powerful model cannot reliably mark a question when the question, context and marking guidance have not been extracted correctly.

I also learned how valuable Codex can be for work beyond writing individual pieces of application code. It helped manage long running processes, investigate failures, improve the data structure and build the infrastructure required to evaluate the system properly.

Most importantly, I learned that the mark itself may not be TensorPaper’s most valuable output.

A mark helps one student understand one answer. Structured responses collected over time could help a teacher understand a class, help a school target support and help an examination board understand where the curriculum is not working as intended.

What is next

The next steps are to:

• Complete and publish the full marking benchmarks
• Evaluate adaptive reasoning across different question types
• Improve uncertainty detection and human review tools
• Build teacher and class analytics dashboards
• Connect misconceptions to targeted revision material
• Add further Computing Science papers and qualification levels
• Use the Codex extraction pipeline to expand into other subjects and examination boards
• Test the platform with students and teachers in real educational settings

TensorPaper is not intended to replace teachers.

It is intended to give students more opportunities to practise, reflect and receive support, while helping teachers focus their limited time where it can make the greatest difference.

The goal is to turn every answer into better feedback for one student and, over time, better evidence for education as a whole.

Built With

• codex
• electron
• llm
• typescript