ExamEval

ExamEval is an AI examiner for handwritten university answer books in Hindi (Devanagari) and English. Indian state universities still evaluate lakhs of handwritten answer books by hand every semester. Results take months, marking is inconsistent between examiners, and re-evaluation requests pile up. I see this problem firsthand while building college software for universities in Madhya Pradesh. A teacher sets up the exam once (questions, model answers, rubric) and uploads scanned PDFs or phone photos of answer books. ExamEval runs a five-step agent pipeline: 1. Prepare pages (PDF/photo normalisation) 2. Segment: SAM 3.1 finds the handwritten answer blocks on each page 3. Read: Muse Spark transcribes the handwriting verbatim (Hindi, English or mixed) and maps each block to its question number 4. Grade: Muse Spark marks each answer criterion by criterion against the rubric, quoting evidence from the student's words 5. Moderate: an independent moderator agent re-checks the original pages and re-marks; disagreement, low legibility or low confidence routes the answer to the teacher Teachers review everything on one screen (page images with detected regions, transcripts, marks, evidence, moderator verdict) and confirm or override in one click. Every model call is visible in an agent trace, and an accuracy panel compares AI marks with teacher marks.