Grading is where the teaching week disappears. Objective sections are quick. Subjective answers are not: each one has to be read, weighed against what the question was actually asking, and given a number that will hold up if a parent queries it three weeks later.
Teachers are also, reasonably, suspicious of software that offers to do this. The suspicion is not about accuracy. It is about who is accountable for the mark. EdunodeX’s answer to that is structural rather than reassuring: the system cannot write a mark on its own.
The stack of answer sheets problem
A class of forty, a paper with eight subjective questions, and three or four minutes of honest attention per answer adds up to most of a weekend. The pressure that creates is well understood by anyone who has taught: later scripts get read faster than earlier ones, and the mark a student receives depends slightly on where their sheet sat in the pile.
Consistency, not speed, is the interesting problem. A system that applies the same rubric to the thirty-eighth script as to the first is doing something a tired human genuinely cannot.
Two separate AI steps, deliberately separated
EdunodeX splits the work into two calls that do not overlap.
The first is transcription. A vision model is given the photograph of the sheet and the list of questions it should be looking for, and it returns each question’s answer text. Its instruction is narrow and explicit: transcribe only. Do not grade, do not correct, do not summarise, do not translate, do not invent. If a question is unanswered or illegible, return an empty string and mark it blank.
The second is grading. The transcribed text — not the image — is graded against the question, its maximum marks, and the model answer or rubric where the teacher supplied one.
Keeping these apart is what makes the output auditable. When a mark looks wrong, a teacher can see whether the model misread the handwriting or misjudged the answer, because the transcription is shown alongside the proposed mark. A single combined step would collapse those two very different failures into one unexplainable number.
Both calls run through EdunodeX’s own AI proxy, so the spend is attributed to the school and the platform retains a per-school cut-off. The feature is gated twice over: the classroom module must be enabled, and a separate AI grading entitlement must be present. Having classroom features is not sufficient, which is how the government-schools package runs classroom workflows with AI grading switched off.
Rubrics, model answers and partial marks
The grading prompt is written for Indian board answer conventions and asks the model to weigh factual accuracy and completeness, correct terminology, logical structure, and relevance to what was actually asked.
Marks can be fractional, because subjective marking is fractional. A five-mark answer that gets the concept right and misses one required example is not a three and not a four.
Where a rubric is supplied, the response scores each criterion separately, so the teacher sees not just the total but its composition. Alongside the mark, the model returns short constructive feedback aimed at the student and a brief statement of why it awarded what it did. The reasoning field is for the teacher; the feedback field is what a student could reasonably be shown.
Confidence scores and the review flag
Every graded question carries a confidence value, and every transcribed answer carries one too. A question whose combined confidence falls below the review threshold is flagged.
This turns a full re-check into a triage list. A teacher scanning a graded sheet does not have to read all eight answers with equal attention; they read the flagged ones properly and spot-check the rest. That is a defensible use of a confidence score, and it is roughly how a moderating examiner already works.
Low confidence typically means one of three things: the handwriting was hard to read, the answer was ambiguous against the rubric, or the answer wandered far enough from the model answer that the model could not tell whether it was wrong or simply different. All three are cases where a teacher should be looking anyway.
Saving a mark is a separate action
The grading endpoint persists nothing. It returns a preview: transcribed answer, awarded marks, maximum marks, feedback, and the review flag, per question.
Saving is a different call. It takes the teacher-reviewed total, validates that the student actually belongs to the class the exam was set for, clamps the value to the exam’s maximum, and upserts the student’s mark for that exam, recording the user who entered it. An absent student can be recorded as absent rather than as a zero, which matters for how averages are later computed.
That separation is the whole argument. The AI proposes a number. A named teacher, on a specific date, decides that number is the student’s mark. If a parent asks later who awarded it, the answer is a person.
Subjects and scripts where this works less well
Mathematics with substantial worked steps is a weak case. Transcription of algebraic and geometric working from a photograph is harder than transcription of prose, and a rubric that awards method marks needs the working read correctly to be applied at all.
Diagrams, maps and labelled figures are outside what this reads. So is anything where the mark depends on presentation rather than content.
Dense or unusual handwriting degrades the whole chain, since grading is applied to whatever transcription produced. This is the failure mode most worth understanding: a confidently graded misreading. The mitigation is that the transcription is always visible next to the mark, so a teacher reviewing a flagged answer sees the text the grade was actually given for.
Finally, answer sheets are student personal data and are handled as such. The photograph is processed to produce a preview; the school decides what is retained, and the mark that lands in the gradebook is the teacher’s, entered under the teacher’s own account.