Accuracy & benchmarks

    How accurately can AI read handwritten answer sheets?

    Chanakya AI reaches ~95% OCR accuracy on real student handwriting and ~97% checking accuracy, with AI-drafted marks typically within 3-5 marks of expert human graders. Here is how the pipeline works and where those numbers come from.

    95%
    OCR accuracy on handwriting
    97%
    Checking accuracy
    3-5
    Marks variance vs. human graders
    <60s
    Per answer, end to end

    From scanned booklet to reviewed marks

    Accuracy is a property of the whole pipeline—not OCR alone.

    Step 1

    Capture

    Teachers photograph or bulk-scan answer booklets. Upload takes about 30 seconds per batch.

    Step 2

    OCR extraction

    Advanced OCR reads handwritten answers across varied handwriting styles and languages, preserving question structure.

    Step 3

    Marking-scheme evaluation

    The extracted text is scored against your marking scheme, with mistake highlights and the correct approach—typically in under 60 seconds per answer.

    Step 4

    Teacher review

    Faculty stay in the loop: every AI draft mark can be reviewed and overridden before it reaches a student.

    Where the numbers come from

    First-party pilot results

    These are results from Chanakya AI school pilots, reported as pilot data. Teachers verified marks throughout.

    Delhi-NCR CBSE pilot

    250+
    Students assessed
    95%
    OCR accuracy
    3-5
    Marks variance

    Classes 9-10 in Mathematics and Science. AI-drafted marks stayed within 3-5 marks of expert human graders, and teachers reported 8+ hours saved per week.

    St. Andrew's ICSE pilot

    200+
    Copies evaluated
    15
    Teachers
    2 mo
    Duration

    A 2-month ICSE pilot across 4 subjects. Granular findings included 95% MCQ accuracy in Class 8A Chemistry while surfacing that 59% of descriptive answers were incomplete—exactly the concept gaps teachers wanted to see.

    Read the full story in the St. Andrew's case study.

    What affects OCR accuracy

    Scan quality drives most real-world accuracy differences. A few habits keep results high:

    • Flat pages with even lighting and the full booklet in frame (roughly 300 DPI equivalent).
    • Consistent booklet orientation so question regions and margins stay readable.
    • English and Hindi handwriting are both supported; extremely faint or overwritten regions are flagged for a human rather than guessed.
    • Diagrams and heavily symbolic work are routed to teachers, since spatial correctness still needs human judgement.

    Accuracy FAQs

    How accurate is Chanakya AI's handwriting OCR?

    In our pilots, handwriting OCR reached ~95% accuracy and overall checking accuracy ~97%, with AI-drafted marks typically within 3-5 marks of expert human graders. Results depend on scan quality and subject.

    Are these third-party benchmarks?

    No. These are first-party results measured during Chanakya AI school pilots (for example the Delhi-NCR CBSE pilot and the St. Andrew's ICSE pilot). We report them as pilot results, and every mark stays reviewable by a teacher.

    What is the best way to verify accuracy for our school?

    Blind-score 30-50 of your own scripts with expert teachers, run the same papers through Chanakya AI, and compare marks and overrides by question type. Accuracy on your handwriting and marking scheme is what matters.