Inspiration

Traditional examination evaluation is often time-consuming, repetitive, and prone to inconsistencies, especially when dealing with large volumes of handwritten answer sheets. Teachers spend hours manually reading, assessing, and grading responses.

We wanted to explore how AI could simplify this process without removing the teacher's judgment. This inspired us to build Ink2Grade, an intelligent platform that makes handwritten answer evaluation faster, more consistent, and transparent.

What it does

Ink2Grade is an AI-powered handwritten answer-sheet evaluation platform that automates the examination assessment workflow.

Key features include:

Handwriting Recognition: Extracts handwritten text from scanned answer sheets using TrOCR.
Rubric-Based Grading: Evaluates answers against teacher-defined rubrics and criteria.
Semantic Evaluation: Understands the meaning of student responses rather than relying solely on keyword matching.
Partial Credit Assessment: Awards marks based on how well individual grading criteria are satisfied.
Human-in-the-Loop Review: Allows teachers to verify uncertain transcriptions and grading decisions.
Audit Trail: Maintains a history of grading decisions, corrections, and overrides.
Multi-Page Support: Processes answers that span multiple pages.
Results & Export: Provides examination results and CSV/JSON exports.

How we built it

We developed Ink2Grade using a modular architecture that combines modern web technologies with AI-powered document processing.

Technology Stack:

Frontend: React, TypeScript
Backend: Python, FastAPI
Database: PostgreSQL
OCR: Microsoft TrOCR for handwritten text recognition
AI/LLM: Semantic evaluation and rubric-based assessment
Infrastructure: Docker and Docker Compose
Storage: Local storage with support for S3 integration

The workflow involves uploading scanned answer sheets, extracting handwritten content, associating responses with questions, evaluating them against predefined rubrics, and presenting results for teacher verification.

Challenges we ran into

Handling variations in handwriting styles, writing quality, and image clarity.
Ensuring accurate text extraction from scanned handwritten documents.
Mapping extracted answers to the correct examination questions.
Designing a grading mechanism that understands semantic meaning rather than depending entirely on exact keyword matches.
Balancing automated evaluation with human oversight to maintain grading reliability.
Building a modular architecture that supports document processing, evaluation, and review workflows.

Accomplishments that we're proud of

Designed an end-to-end architecture for automated handwritten answer-sheet evaluation.
Integrated TrOCR for handwritten text recognition.
Developed a rubric-driven evaluation approach that supports partial credit.
Incorporated human-in-the-loop verification to keep teachers involved in critical decisions.
Designed an audit trail to improve transparency and accountability.
Built a scalable foundation using FastAPI, React, PostgreSQL, and Docker.

What we learned

Through Ink2Grade, we explored the practical challenges of combining computer vision, natural language processing, and generative AI in an educational application.

We learned that building a reliable AI system requires more than model integration. Data quality, evaluation consistency, explainability, confidence assessment, and human oversight are equally important.

We also gained experience in designing modular backend architectures, integrating AI services, and developing workflows that address real-world educational challenges.