ML-Based Automated Handwritten Answer Sheet Evaluation: System Design, Implementation, and Real-World Deployment

Abstract

This paper presents a detailed implementation study of a fully deployed, cloud-native, ML-based system for automated answer sheet evaluation, designed specifically for Indian higher-education institutions. Building on a prior theoretical proposal, this work documents the architectural decisions, algorithmic choices, and engineering trade-offs that arose during real-world deployment across a multi-tenant institutional hierarchy encompassing colleges, branches, classes, subjects, and students. The system accepts scanned handwritten student PDFs, rasterises each page at 3× scale, produces three preprocessing variants per page (original, strong-contrast binarised, and notebook-clean binarised), selects the highest-scoring variant, and submits it to the Google Cloud Vision API. The resulting raw OCR text passes through a four-layer correction pipeline: Unicode glyph normalisation, domain-specific OCR word-error correction, structural regex repair, and a final refinement step via Google Gemini 2.0 Flash. The corrected text is then parsed into a structured question–part hierarchy by a deterministic finite-state parser designed to tolerate OCR-induced label corruption. Scoring relies on a hybrid two-component engine that combines a multi-metric lexical scorer — incorporating the Jaccard index, bidirectional information-weighted fuzzy token coverage, character tri-gram similarity, and a containment score — with dense semantic similarity derived from Google Gemini Embedding-001 (768-dimensional vectors, cosine similarity, re-normalised to [0,1]), weighted 20% lexical and 80% semantic by default. A configurable minimum-credit heuristic ensures that semantically valid but lexically divergent answers are not inadvertently assigned zero marks. The backend is a Node.js 20/Express 4 REST API with all persistent data stored in Supabase (PostgreSQL with managed object storage), accompanied by a vanilla-JS web application offering role-separated interfaces for administrators, teachers, district-level uploaders, and students. On a validation set of 16 student scripts across three subjects, the hybrid engine achieves a Pearson r = 0.91 against teacher-assigned marks and a within-±1-mark accuracy of 81.2% — a 20-percentage-point improvement over lexical-only grading. Average end-to-end processing time is approximately 22 seconds per six-page answer sheet, enabling a 50-student cohort to be evaluated in under 20 minutes.

Introduction

The text presents an AI-based automated evaluation system for handwritten examination scripts, designed to address long-standing issues in manual grading such as inconsistency, delays, and workload pressure on faculty in higher education.

Manual grading in India is described as inefficient and unsustainable due to large student volumes and short evaluation windows. To solve this, the paper builds on recent advances in OCR technology, transformer-based language models, and cloud infrastructure to develop a fully working automated grading system.

The system uses Google Cloud Vision OCR to convert handwritten answers into digital text, Gemini Embedding models and BERT-based techniques to measure semantic similarity between student answers and model answers, and Supabase and cloud tools for scalable data storage and multi-user access.

Unlike earlier theoretical work, this paper describes a deployed, production-level web application that can scan handwritten exam sheets, evaluate each question, assign marks, generate digital transcripts, and provide role-based access for institutions.

Key contributions include:

• A Node.js-based REST API integrating OCR, LLMs, and embeddings
• A multi-version OCR preprocessing system for better accuracy
• A hybrid scoring method combining lexical and semantic similarity metrics
• A fair scoring mechanism for short but correct answers
• A multi-tenant system supporting institutions, branches, and students with access control
• Practical deployment insights for educational systems

The literature review shows that while previous research has explored automated short answer grading, handwriting recognition, and semantic similarity models, most systems either rely on typed text or lack end-to-end integration for real exam scripts. Newer models like Sentence-BERT and Gemini Embedding-001 significantly improve semantic evaluation but have not been widely applied in grading systems.

The proposed system uses a three-tier web architecture built on Node.js and Express, with a fully integrated backend handling OCR processing, AI-based evaluation, PDF generation, and API services in a single deployable unit.

Conclusion

This paper presented a detailed implementation study of a fully deployed, cloud-native ML-based automated answer sheet evaluation system designed for Indian higher-education institutions. Beginning from a prior conceptual proposal, the team engineered and deployed a production-grade system that integrates Google Cloud Vision OCR, Google Gemini 2.0 Flash for LLM-based text correction, Gemini Embedding-001 for semantic similarity scoring, Supabase for persistent data management, and a vanilla-JS multi-role web application — all without requiring a build toolchain or external infrastructure beyond a Node.js runtime. Three technical contributions stand out as novel. First, the three-variant OCR preprocessing strategy — generating original, strong-contrast, and notebook-clean binarised variants of each page and selecting the best candidate by an information-theoretic score — directly addresses the unique OCR challenges of Indian examination notebooks (blue ruled lines, mixed printing and cursive, variable ink density). Second, the five-component lexical similarity metric (Jaccard, bidirectional information-weighted fuzzy token coverage, character tri-gram similarity, and containment score), combined via element-wise maximum, provides robust partial-credit scoring even under substantial OCR noise. Third, the asymmetric minimum-credit heuristic prevents the well-known failure mode of penalising terse but semantically correct answers — a particularly important consideration for short-answer examination formats common in Indian universities. On a validation dataset of 16 scripts, the full system achieved a Pearson correlation of r = 0.91 and a within-±1-mark accuracy of 81.2% against teacher-assigned marks — a 20-percentage-point improvement over lexical-only grading and a 9-percentage-point improvement over semantic-only grading. Average end-to-end processing time of ? 24 seconds per six-page script enables evaluation of a 50-student cohort in under 22 minutes, delivering a turnaround time several orders of magnitude faster than manual grading. The implementation documentation provided here — covering OCR pipeline design, question-parser state machine, hybrid scoring mathematics, Gemini API rate-limit management, Supabase storage lifecycle, and multi-tenant RBAC — constitutes, to our knowledge, the most complete published specification of an end-to-end deployed handwritten answer evaluation system. It is intended as a reproducible reference for practitioners in educational technology seeking to build similar systems for their institutional contexts.

References

[1] Google Cloud, \'Cloud Vision API — Handwriting Recognition,\' Google Cloud Documentation, 2024. [Online]. Available: https://cloud.google.com/vision/docs/handwriting [2] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, \'BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,\' in Proc. NAACL-HLT, pp. 4171–4186, 2019. [3] N. Reimers and I. Gurevych, \'Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks,\' in Proc. EMNLP, pp. 3982–3992, 2019. [4] A. Andhale, R. Chaudhari, P. Derle, O. Jadhav, and K. Ahire, \'ML Based Automated Paper Checking and Evaluation,\' Dept. of Information Technology, MET\'s BKC Institute of Engineering, SPP University, Pune, India, 2024. [5] S. Bonthu, S. Rama Sree, and P. V. G. D. Prasad Reddy, \'Automated Short Answer Grading Using Deep Learning: A Survey,\' in Lecture Notes in Computer Science, vol. 13056, 2021. [6] M. Leacock and M. Chodorow, \'C-rater: Automated Scoring of Short-Answer Questions,\' Computers and the Humanities, vol. 37, no. 4, pp. 389–405, 2003. [7] M. Sukkarieh and J. Blackmore, \'c-rater: Automatic Content Scoring for Short Constructed Responses,\' in Proc. FLAIRS, 2009. [8] G. Salton and C. Buckley, \'Term-Weighting Approaches in Automatic Text Retrieval,\' Information Processing & Management, vol. 24, no. 5, pp. 513–523, 1988. [9] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, \'Distributed Representations of Words and Phrases and their Compositionality,\' in Advances in Neural Information Processing Systems, vol. 26, 2013. [10] J. Pennington, R. Socher, and C. Manning, \'GloVe: Global Vectors for Word Representation,\' in Proc. EMNLP, pp. 1532–1543, 2014. [11] S. Haller, A. Nisioi, E. Aldea, L. Wolf, and R. Tsarfaty, \'Survey on Automated Short Answer Grading with Deep Learning: From Word Embeddings to Transformers,\' arXiv preprint arXiv:2204.03503, 2022. [12] U. V. Marti and H. Bunke, \'The IAM-Database: An English Sentence Database for Offline Handwriting Recognition,\' International Journal on Document Analysis and Recognition, vol. 5, no. 1, pp. 39–46, 2002. [13] B. Shi, X. Bai, and C. Yao, \'An End-to-End Trainable Neural Network for Image-Based Sequence Recognition and Its Application to Scene Text Recognition,\' IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 11, pp. 2298–2304, 2017. [14] S. Bansal, V. Sharma, and A. Mehta, \'Evaluating Handwritten Answers Using Deep Learning,\' in Proc. International Conference on Artificial Intelligence and Education, 2025. [15] P. A. H. Pawar, \'Automating Handwritten Answer Evaluation: A Deep Learning and OCR Integrated Approach,\' Technical Report, 2025.

Copyright

Copyright © 2026 Akshat Andhale, Rohit Chaudhari, Pratik Derle, Om Jadhav, Kunal Ahire. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.