Transforming Student Answer Sheets into Actionable Classroom Intelligence
ClassAssist AI was awarded as one of the Top 3 Best Projects among 14 finalist teams in the LLM/SLM Contest conducted at COEP Technological University.
Evaluating answer sheets manually is both time-consuming and limited in the insights it provides.
Teachers often spend hours checking answer sheets, yet marks alone cannot explain:
- Why students lost marks
- Which concepts the class struggles with
- What topics require revision
- How teaching strategies should be adapted
ClassAssist AI transforms raw answer sheets into meaningful classroom intelligence within minutes.
Unlike conventional grading systems, ClassAssist AI focuses on conceptual understanding rather than marks.
Instead of simply reporting scores, it identifies recurring misconceptions across an entire classroom, groups similar learning gaps using semantic clustering, and assists teachers in planning targeted remedial instruction.
For a class of 50 students:
- Manual evaluation: ~9 hours
- ClassAssist AI: ~1 hour (current laptop implementation)
≈ 9× faster classroom insight generation
- Upload multiple student answer sheets
- Upload a single teacher answer key
- Batch analysis of an entire classroom
- Sarvam AI OCR
- PDF, PNG, JPG & JPEG support
- Preserves document structure
Each student's answers are compared against the teacher's answer key using a locally deployed Qwen 2.5 3B model via LM Studio.
The system identifies:
- Topics covered
- Conceptual misconceptions
- Learning strengths
- Learning weaknesses
- Overall conceptual understanding
Instead of generating duplicate observations like:
- Student 1 → Confuses Force & Momentum
- Student 2 → Mixes up Momentum and Force
- Student 3 → Treats Momentum as a Force
ClassAssist AI embeds these observations using Sentence Transformers and groups semantically similar insights into a single classroom misconception.
Teachers receive class-level patterns instead of repetitive student-wise feedback.
- Topics identified
- Common misconceptions
- Class strengths
- Learning weaknesses
- Topic-wise insights
- Severity analysis
Automatically generates:
- Classroom performance summary
- Priority topics for revision
- Lesson objectives
- Classroom activities
- Actionable teaching recommendations
Generates a downloadable report containing:
- Topics
- Misconceptions
- Strengths
- Weaknesses
- Teaching Plan
- Teacher Summary
Student Answer Sheets Teacher Answer Key
│ │
└──────────────┬─────────────┘
▼
Sarvam AI OCR
▼
Structured Text Extraction
▼
Qwen 2.5 LLM Analysis
▼
Semantic Clustering
▼
Classroom Analytics Engine
┌─────────┴─────────┐
▼ ▼
Interactive Dashboard PDF Report
│
▼
AI-Generated Teaching Plan
| Component | Technology |
|---|---|
| OCR | Sarvam AI OCR |
| LLM | Qwen 2.5 3B via LM Studio |
| Embeddings | all-MiniLM-L6-v2 (locally stored) |
| Dashboard | Streamlit |
| Backend | Python |
| ReportLab | |
| ML Libraries | Sentence Transformers, Scikit-Learn |
- Upload student answer sheets
- Upload teacher answer key
- OCR extracts text
- Qwen analyzes conceptual understanding
- Semantic clustering groups similar insights
- Classroom analytics are generated
- AI creates a teaching plan
- Dashboard and PDF report are generated
git clone https://github.com/ryzenBot72/Class-Assist-AI.git
cd Class-Assist-AI
pip install -r requirements.txtConfigure your .env file:
SARVAM_API_KEY=YOUR_API_KEYClassAssist AI uses local models for its analysis pipeline:
- Qwen 2.5 3B is hosted and run locally through LM Studio.
- all-MiniLM-L6-v2 is stored locally on the machine and loaded directly during semantic clustering.
Before running the application, ensure that the Qwen model is available and running through LM Studio with the required local API endpoint configured.
The semantic embedding model should also be present in the expected local model directory.
Note: The embedding model is not included in this repository due to its size. It must be downloaded separately and placed in the expected local model directory before running the application.
streamlit run app.py- Batch processing of multiple answer sheets
- AI-powered conceptual understanding analysis
- Semantic clustering of misconceptions
- Classroom-level learning analytics
- Interactive dashboard
- AI-generated teaching plans
- Professional PDF reports
The dashboard provides a class-level overview of identified topics, common misconceptions, severity, strengths, and weaknesses.
 |
 |
The system generates a structured teaching plan based on the identified classroom learning gaps.
The generated report consolidates classroom insights, misconceptions, strengths, weaknesses, and the AI-generated teaching plan.
 |
 |
 |
 |
- Student dashboard
- Retrieval-Augmented Generation (RAG) to remove dependency on answer keys
- Support for multiple school boards
- Subject-specific fine-tuned models
- Cloud deployment for schools
- Sarvam AI for OCR capabilities
- Hugging Face for transformer models
- Qwen Team for the open-source LLM
- COEP Technological University for organizing the LLM/SLM Contest
- Aditi Bhise
- Manthan Bhala
- Sangram Lohokare
This project is licensed under the MIT License.