AI-Powered Question Solving & Evaluation Platform

Overview

Instead of forcing image questions through one method, the platform selects among OCR, multimodal vision, and hybrid processing and presents results in a comparable evaluation flow.

Problem and goal

Text-heavy, formula-based, and visually contextual questions need different preprocessing. One pipeline does not behave equally well for every question type.

My role

• Unified OCR, vision, and hybrid routes behind one service contract.
• Developed image preprocessing and adaptive selection logic.
• Split pipeline behavior into testable units with Pytest.

Solution

• Input quality and question structure informed pipeline selection.
• OpenCV/Tesseract handled extraction while a multimodal LLM supported visually contextual questions.
• A Streamlit interface made method results easier to review.

Architecture

The system’s main components and data/decision flow are separated as follows.

Tech stack

Implementation decisions

• Use a shared result contract for OCR, Vision, and Hybrid outputs.
• Base adaptive routing on OCR signals and visual question properties.

Results and evaluation

• Compared OCR, Vision, Hybrid, and Adaptive pipelines in one evaluation framework.
• The repository reports 140 passing Pytest tests; Adaptive scored 31/34 on the expanded synthetic set and 11/12 on the realistic synthetic set, while Hybrid scored 10/12.

Challenges and solutions

• Variable image quality → preprocessing plus alternative pipelines.
• Different output shapes → a shared response and evaluation contract.

What I learned

The work showed the importance of input analysis before model calls and of testable pipeline boundaries in adaptive AI applications.