Problem & context
Problem
Convert raw mathematics-assessment exports into privacy-conscious, analysable records and help researchers interpret recurring student errors at scale.
Intended users
Mathematics education researchers and educators preparing and analysing student-response datasets.
Role and contributions
Workflow design, data processing and AI-assisted error-analysis implementation
Dataset / inputs
Paired grade and response exports, student answers, reference answers and optional mathematical context. The batch endpoint processes up to 50 valid rows per request.
Methods & architecture
The workflow assigns pseudonymous student identifiers, removes technical markers, normalises grades, filters unusable or fully correct rows when requested, and applies a three-stage language-model chain for explanation, Newman classification and concise summary.
Implementation
- Anonymise student names into stable numeric identifiers
- Clean STACK response markers and normalise question grades
- Filter blank, invalid and explicitly no-error records
- Analyse rows with bounded concurrency and per-row failure isolation
- Support OpenAI and OpenRouter-assisted batch analysis
- Export prepared and analysed data for downstream research
Current evidence
The current implementation verifies the project foundation below; broader outcome claims will be added after evaluation.
Limitations & next steps
The local application contains no automated test suite, and its production build was not fully confirmed during this review. AI-generated explanations and error categories require human validation. The GitHub repository was not publicly accessible at review time and is behind the local code.
Add a labelled evaluation set, inter-rater validation, automated tests and a reproducible benchmark before reporting classification quality.
Attribution & licence
No licence file was verified in the local repository.