The Problem

Grading at scale is slow, inconsistent, and mentally exhausting. Teachers managing large classrooms found themselves applying marking criteria differently day to day the same answer scored differently depending on fatigue, time pressure, or context drift.

Existing AI grading tools forced a binary choice: save time or stay accurate. Opaque scores without reasoning made it impossible for teachers to trust or challenge the output. The real problem wasn't automation, it was the absence of accountability.

The Process

Using the Double Diamond framework, the core design challenge was reframed: not "how do we automate grading?" but "how do we make AI-assisted grading something teachers can actually trust and own?"

Research led to two defining decisions. First, the AI was given three distinct strictness personas calibrated like a grading tone so teachers could set expectations that matched the context (formative quiz vs. high-stakes assessment). This translated a complex technical variable into a human-readable choice.

Second, the team chose to grade one submission at a time rather than in bulk. Speed was deliberately traded for grounding every score surfaced criteria-wise reasoning, giving teachers the transparency to agree, override, or annotate with their own judgment.

The Outcome

The feature shipped live, fully integrated into the existing student submission platform no new workflow, no new tool to learn. Teachers grade inside the same environment where students already submit.

The result was measurable and immediate: teachers reported saving ~60% of the time previously spent on manual grading each week time redirected toward instruction, intervention, and student support.

The system earns trust through four qualities: transparent reasoning per criterion, full grade override control, adaptive learning from teacher feedback, and single-submission focus that keeps evaluation intentional.