Inspiration
Every year, educators spend hundreds of hours manually grading assignments, leaving them exhausted and with little time to focus on personalized student mentorship. While multiple-choice digital assessments exist, they fail to capture a student's true critical thinking process. Real education happens on paper—through handwritten equations, scratch work, and structured sentences.
We were inspired to build Kagaz-AI to bridge the gap between traditional pen-and-paper learning and modern classroom analytics. By automating the tedious process of review while deeply analyzing how a student arrived at an answer, we unlock macro classroom intelligence that helps teachers catch learning gaps before they widen.
What it does
Kagaz-AI is an automated educational evaluation platform that transforms handwritten student worksheets into deep pedagogical insights.
- Asynchronous Grading: Teachers can snap a photo or batch-upload student worksheets. The system processes the pages in the background, minimizing teacher wait times.
- Pedagogical Gap Tracking: Instead of just marking an answer right or wrong, the platform breaks down why a student made an error, tracking concept-level understanding over time.
- Macro Classroom Intelligence: Aggregated student metrics flow seamlessly into an interactive teacher dashboard, highlighting which topics require immediate reinforcement across the entire classroom.
How we built it
We designed Kagaz-AI with a production-grade, decoupled full-stack architecture to handle intensive AI processing workflows efficiently:
- Frontend: Built with Next.js and styled with Tailwind CSS, providing an intuitive, highly responsive dashboard for tracking batches, uploading sheets, and visualizing classroom analytics.
- Backend Backend API: Powered by FastAPI, serving high-performance, asynchronous endpoints for lightning-fast request routing.
- Task Queue & Orchestration: To prevent heavy AI computation from locking the main web server, we implemented a decoupled task architecture using Celery. This offloads background image extraction and evaluation workflows safely.
- AI Evaluation Logic: The platform handles multimodal verification by pairing customized OCR parsing routines with structured vision-LLM logic. It converts raw handwriting pixels into precise JSON structures evaluating performance data.
Challenges we ran into
Building a reliable, automated grader introduces unique engineering constraints:
- Handling Heavy Multimodal Pipelines Asynchronously: Running intensive vision tasks over multiple student papers simultaneously risked crashing the server. We solved this by strictly segregating our network request handling from computational tasks via Celery background workers.
- Noisy Handwriting & OCR Hallucinations: Real-world student sheets are rarely pristine. Dealing with varied handwriting sizes, angled photos, and stray marks meant our processing prompts had to be heavily engineered with explicit structural formats to enforce consistent evaluation criteria.
- Structured JSON Output Guarantees: For data to render on the Next.js dashboard correctly, the AI backend must always return perfect schema compliance. We addressed this by implementing strict JSON parsing constraints on the processing model's returns.
Accomplishments that we're proud of
- End-to-End Async Automation: Achieving a completely non-blocking UI experience where a teacher can dump a massive file of papers and browse analytics instantly while backend workers compute in the background.
- High Structural Fidelity: Designing an API structure capable of mapping handwritten visual mistakes accurately to concept-level performance markers.
- Clean, Production-Ready Modular Stack: Crafting a clean codebase separation where frontend interfaces and heavy task processors communicate reliably with zero overlap.
What we learned
- The Reality of Asynchronous Architecture: We gained hands-on experience structuring reliable task messaging, dealing with distributed states, and designing reliable fallback statuses when background workflows execute.
- The Importance of Schema Validation: We learned that when interfacing AI with modern frontends, robust validation layers are non-negotiable for protecting the UX against unstructured machine outputs.
- Context-Driven Prompts Over Generalities: We realized that making AI a helpful educational tool requires highly specialized instruction constraints rather than standard out-of-the-box system configurations.
What's next for Kagaz AI
Kagaz-AI has a clear operational roadmap to move from a prototype to a comprehensive school platform:
- Localized Language Support: Enhancing OCR capabilities to process regional languages and dialects, making the platform accessible to underfunded or multilingual public school systems.
- Formula Performance Tracking: Integrating specific token configurations to handle advanced multi-line mathematical proofs seamlessly. Let a student's derivation mistake be mapped by the model using formal evaluation logic, satisfying expressions like:
$$\Delta y = f'(x)\Delta x + \epsilon\Delta x$$
- Automated Remedial Generator: Utilizing macro gap tracking metrics to automatically generate custom review sheets targeted uniquely to individual student struggles, completely automating personalized remedial support.
Log in or sign up for Devpost to join the conversation.