J3

Inspiration

Every few years, the national exam board updates their syllabi, making many questions now obsolete. As tutors, having to cross reference old exam papers against the new curriculum is tedious and steals hours from what matter most, teaching.

What it does

Our platform acts as an AI-powered Curriculum Auditor

1. Compare Syllabi Compares 2 syllabus PDFs, identify what has been added, removed and modified and present it to the user. Students planning to or currently tutoring younger students may find this feature helpful to find out how the syllabus has changed since they graduated and what they should focus on when teaching younger students.

2. Paper Alignment Takes in a syllabus PDF and a practice paper, identifies topics of each question and determine whether the question is still aligned to the topics mentioned in the syllabus PDF. Students that are acting as tutors may find this feature helpful in analysing how useful and relevant their past papers are for their current tutee based on the current syllabus.

3. Check Mapping Compares 2 syllabus PDFs, map topics between the 2 syllabus PDFs and determine similarity. Students planning for their exchange programmes may find this feature helpful to compare course outlines of their university and their exchange university to determine whether the courses are similar enough to be mapped.

How we built it

PyMuPDF (fitz) in a FastAPI backend to parse raw text from unstructured PDFs.

The raw text is fed into our AI model, which uses strict JSON schemas to transform messy descriptions.

We implemented a semantic comparison logic where the AI acts as an auditor, comparing two JSON structures to calculate Similarity Scores and identify content gaps.

The Next.js frontend consumes these JSON reports, rendering them into interactive comparison tables and Recharts visualizations for the user.

Challenges we ran into

Large documents can be dozens of pages long. We had to implement a chunking strategy to feed relevant sections of the curriculum into the LLM without hitting token limits.

Building without a database meant needed clean JSON schemas to ensure the UI could instantly render complex comparison reports from a single API call.

We realised that simply uploading a PDF to the AI didn't work. summarise too much, miss subtle syllabus changes, or hallucinate that two modules were identical just because they had similar names.

Accomplishments that we're proud of

One of our key accomplishments is the frontend design, where we focused on creating a clean, intuitive, and user-friendly interface that makes the workflow easy for tutors to understand and use with minimal learning curve. We are also proud of how we seamlessly integrated all components of the system, from input handling and prompt orchestration to result generation into a smooth user experience. can u say sth else

What we learned

Through this project, we learned the importance of crafting well-designed prompts to guide the model towards generating accurate and relevant outputs. Small changes in prompt structure, constraints, and examples had a significant impact on output quality.

We also learned the critical role of input preprocessing and cleaning. By filtering out noise, irrelevant information, and inconsistencies from the input data, we were able to greatly improve the reliability and consistency of the generated results.

What's next for J3

Moving forward, we plan to extend the platform beyond a single subject to support multiple academic disciplines, making it a versatile tool for tutors teaching different subjects.

In addition, we aim to integrate a backend system that allows tutors to store, manage, and retrieve syllabi seamlessly. This will eliminate the need for repeated manual uploads and enable more efficient syllabus comparisons over time. Tutors will be able to reuse previously saved syllabi, track changes, and perform comparisons with minimal friction.

Beyond syllabus comparison, we also plan to expand the platform’s capabilities to question generation. By leveraging past-year papers and syllabus alignment, the system will be able to generate similar practice questions, helping tutors quickly create revision materials tailored to their students’ needs. This feature will significantly reduce preparation time while maintaining quality of questions.

Built With

• fastapi
• nextjs
• openai
• pymupdf
• python
• react-dropzone
• recharts
• tailwind
• uvicorn