AI-Assisted Workflow for Grading Scanned Handwritten STEM Exams
Overview
AI-assisted, human-supervised workflow for grading scanned handwritten STEM exams (including diagrams). Designed for universities and schools that already use paper exams and scanning, but also for commercial education/training providers which struggle to obtain workforce for manual grading. It reduces repetitive grading work while keeping full instructor control: the system produces draft scores and feedback, then highlights borderline or inconsistent cases for quick human review rather than replacing the grader. The workflow decomposes each exam into per-question tasks, applies layout/segmentation preprocessing, and runs multimodal grading with strict templates. Guardrails check answer presence and enforce machine-readable outputs, while optional ensemble passes plus a supervisor step improve robustness. Compute-heavy stages are HPC-ready; development can use hosted model APIs, with a migration path to open-source on-prem models. No foundation-model training is performed.
Outcome
Outputs include structured grading reports (Markdown/JSON), XLSX gradebooks, per-student PDFs, and a traceable manual-override mechanism. The service is batch-oriented and can process large exam campaigns asynchronously via queued jobs, providing progress tracking and reproducible artefacts per run.
Service Card
| Category | AI products |
| Domain | Education, Assessment, Training |
| Vertical | Digital Society |
| Target users | Universities and Schools, Exam Coordinators, Teaching Staff, Commercial Training Providers |
| Service type | Pipeline |
| Outputs | Per-student grading reports (structured text + optional PDFs), aggregated grade spreadsheets (XLSX), audit artefacts (JSON/structured logs), a structured manual-review/override mechanism, and optional task decomposition metadata (per-question mapping/crops). |
| Requirements | Requires scanned exam PDFs/images and a short set of grading rules; best results when a reference solution is provided. Batch-oriented (not real-time) and designed for human-supervised use: the workflow flags uncertain or inconsistent cases for manual review and supports explicit instructor override. Early-phase execution may rely on external hosted multimodal models until AI Factory on-prem capacity is available; migration targets open-source models deployable on SLAIF infrastructure. |
| Limitations | |
| Availability | Q2/2027 |
| Licensing | TBD (workflow code intended for open distribution where possible after large-scale validation; model licences depend on selected backends) |
| Visual content | |
| Visual caption |