Open-source grading system combines local OMR with multimodal AI
A developer presented an open-source answer-sheet grading system that uses local OMR for multiple-choice questions and a single multimodal model request for fill-in-the-blank and essay questions. It supports PDF or image templates, whole-class batch processing, teacher review, CSV exports, and local deployment using React, TypeScript, Tailwind, FastAPI, and SQLite. The developer reports roughly 99% multiple-choice recognition accuracy on one displayed answer sheet, where 3 of 66 questions were flagged for review; this was not independently validated and does not represent overall grading accuracy.
The project assigns deterministic multiple-choice scoring to local OMR, subjective transcription and scoring to a multimodal model, and final decisions to teachers through a review workflow before grades are published.
技术栈：React + TypeScript + Tailwind / FastAPI + SQLite，前后端同仓，本地部署，数据不出校。
一、先看效果：从"通宵批卷"到"几分钟出一个班"
不铺垫，先看这套系统跑起来是什么样子，再讲为什么这么做。
登录后是工作台，一条完整的批改流水线被组织成左侧的几个模块：
每个模块的职责一句话讲清：
#Open-source answer-sheet grading system#OMR#Multimodal LLM#React#TypeScript#Tailwind#FastAPI#SQLite