Essay Grading: Automated Assessment on GPU

The Challenge: 12,000 Essays and a 21-Day Wait for Feedback

A Russell Group university’s business school processes approximately 12,000 undergraduate essays across three assessment windows each academic year. With 45 academic staff sharing marking duties alongside research and teaching commitments, the average turnaround from submission to marked feedback is 21 days — well beyond the university’s 15-working-day target. Late feedback means students cannot incorporate lessons into subsequent assignments. The quality assurance process requires second marking of 20% of submissions and external examiner review, adding further delays. Student satisfaction surveys consistently flag marking turnaround as the school’s weakest metric.

The school needs AI-assisted first-pass assessment that provides students with immediate formative feedback on submission, while generating draft grades and structured commentary for academic staff to review, adjust, and finalise. Student work is confidential academic data — essays contain student IDs, sometimes personal reflections, and represent intellectual property. Routing this through external AI APIs is prohibited by the university’s data protection policy.

AI Solution: LLM-Powered Essay Assessment

A fine-tuned open-source LLM processes each submitted essay against the module’s marking rubric, generating: a proposed grade band, structured feedback across rubric criteria (argumentation, evidence use, critical analysis, structure, referencing), and specific improvement suggestions. The model is trained on 8,000 previously marked essays paired with their grades and marker commentary, learning the school’s assessment standards and feedback conventions.

Running on a dedicated GPU server with vLLM, the system processes an entire 4,000-essay cohort submission overnight. Students receive formative AI feedback within hours of submission. Academic markers then review the AI’s draft assessment alongside the essay, adjusting grades and feedback as needed — a process that takes 5-8 minutes per essay rather than the current 20-25 minutes for first-pass marking from scratch.

GPU Requirements

Essay assessment requires processing long-context inputs (2,000-5,000 word essays) and generating detailed outputs (500-800 word feedback per essay). A model with strong reasoning capabilities — 13B parameters or larger — delivers the assessment quality needed for academic credibility.

For overnight batch processing of a full cohort, the RTX 6000 Pro completes the job in 18 hours. For schools with smaller cohorts or staggered submissions, an RTX 6000 Pro provides excellent value. Private AI hosting ensures all student work remains within UK infrastructure.

Recommended Stack

• vLLM for efficient long-context inference with continuous batching.
• LLaMA 3 70B or Mixtral 8x7B (quantised) fine-tuned on the school’s marked essay corpus using LoRA.
• LangChain for structured prompt engineering, injecting rubric criteria and grade descriptors into each assessment prompt.
• PostgreSQL for storing assessment results, linking AI feedback to student submissions and module specifications.
• VLE integration (Moodle, Blackboard, Canvas) via API for automated submission ingestion and feedback delivery.

For handling handwritten exam scripts, add PaddleOCR or document AI to digitise handwritten responses. Deploy a vision model for assessing diagram quality in technical subjects.

Cost Analysis

The current marking process consumes approximately 5,000 academic staff hours per assessment period across the school. At average academic salary costs, this represents roughly £200,000 per assessment window. AI first-pass assessment reduces the human marking time per essay by 60%, freeing approximately 3,000 hours per window for research and teaching. Over three assessment windows annually, the productivity recapture exceeds £360,000.

The student experience improvement is equally significant. Receiving formative feedback within hours rather than weeks enables iterative learning. The school targets a 10-percentage-point improvement in NSS assessment and feedback scores, directly impacting league table positioning and prospective student recruitment.

Getting Started

Export 5,000 marked essays with grades and feedback from your VLE, anonymising student identifiers. Fine-tune the model on this dataset, then validate against 500 held-out essays, comparing AI grades against human grades. Target less than 5% grade deviation (within one grade band) before deploying to students. Run in shadow mode for one assessment cycle — generating AI feedback internally but not releasing it to students — while academics calibrate their trust in the system.

GigaGPU provides UK-based dedicated GPU servers for education AI workloads. Add an AI chatbot for student assignment queries, or scale GPU allocation during peak assessment periods.