Academic paper correction
Evaluate handwritten and typed answer sheets in batches. The model proposes marks and feedback for every question, and a human reviewer adjusts and finalizes each result.
AI-assisted evaluation of written and subjective answers against the question paper and marking scheme, with reviewer override and a locked final result.
- Images
- Typed and handwritten sheets
- OCR and layout extraction
- Model-agnostic language models
- Structured JSON output
- Rubric-driven prompts
- Background job queue
- Batch tracking
- Cached question paper reading
The challenge
Marking written and subjective answers is slow, repetitive work. Examiners read hundreds of sheets against the same paper, and tired eyes drift. Two markers can score the same answer differently, and students rarely see why they lost marks.
Handing the job to a model without control creates a different risk. Unexplained scores, no way to correct them and no record of who changed what will not stand up to a student appeal or an audit. Institutions need speed that stays accountable.
What it does
Batch answer sheet intake
Select many handwritten or typed answer sheets in PDF, image or document form and submit them as one batch, grouped together for tracking and review.
Question paper and marking scheme
Supply the question paper and an optional marking scheme or model answers, as a file or as pasted text, so grading follows the examiner's own criteria.
Per-question scoring
Each question receives marks awarded, maximum marks, a status of correct, partial or incorrect, and written feedback that names what was right and what was missing.
Partial credit rules
Partial credit scales with the marks at stake. Method marks are given for correct working, and unanswered or unreadable answers are flagged instead of guessed.
Reviewer override
A reviewer sees the answer sheet, the question paper and the model's results side by side, overrides any mark and adds a comment before saving.
Finalize and export
Finalizing locks a reviewed sheet against further edits. Results export to CSV with one row per question, and batch views show averages and status.
How it works
Configure the exam
Choose the model deployment, then provide the question paper and, optionally, the marking scheme or standard answers as files or text.
Submit the batch
Pick answer sheets from storage and start the evaluation. Sheets are queued and processed in turn, and progress refreshes on screen.
Read and grade
An OCR engine reads each sheet, matches answers to questions, and the model scores each one against the paper and scheme with structured feedback.
Review and finalize
A reviewer checks results beside the original sheet, overrides marks where needed, finalizes the sheet, and exports the result as CSV.
Under the hood
- Document intake
- Images
- Typed and handwritten sheets
- OCR and layout extraction
- Grading
- Model-agnostic language models
- Structured JSON output
- Rubric-driven prompts
- Workflow
- Background job queue
- Batch tracking
- Cached question paper reading
- Output and storage
- CSV export
- JSON results
- Cloud object storage
- REST
Business outcome
Faster turnaround on large batches. Examiners start from a scored, explained draft for every sheet instead of a blank page, so their time goes to judgement calls.
More consistent marking. The same paper, scheme and rules apply to every sheet in a batch, which reduces drift between markers and between sessions.
Feedback students can use. Each question carries specific feedback and an improvement tip, so a result explains itself and supports learning, not only grading.
Human accountability kept. Reviewers hold the final say. Overrides, comments and locked results give the institution a defensible record for any dispute.
Where it is used
- A school board marks a term examination, with teachers reviewing and finalizing the model's proposed scores.
- A university department grades short-answer and essay questions across a large cohort using one shared marking scheme.
- A training provider scores written assessments and returns question-level feedback to learners.
- A certification body reviews flagged or illegible answers while routine sheets move through the batch.
Frequently asked questions
Does the AI decide the final mark?
No. The model proposes marks and feedback. A reviewer can override any mark with a comment, and only the reviewer finalizes a sheet. Finalized sheets are locked against further edits.
Can it read handwriting?
It uses OCR and layout extraction to read handwritten and typed sheets. Answers that cannot be read are flagged for human review rather than scored, because accuracy depends on scan quality.
Which AI model does it use?
You choose the model deployment for grading. The product is model-agnostic, so frontier, open-weight or small language models can be used, subject to your policy.
Do we need a marking scheme?
It is optional. A scheme or model answers improve consistency. Without one, the model uses the question paper and applies conservative judgement, which reviewers should check more closely.
See it on your own documents and data
Tell us what you want to automate and we will show you how it runs inside the harness.