Automatic job correction machine

Login page
Data statistics
Project Homepage
Technical architecture diagram
Product Introduction

Inspiration

In the educational setting, teachers have long been burdened with the repetitive task of grading homework, and traditional tools (such as OCR scanning + basic algorithms) have low accuracy (only 95%) and insufficient software and hardware integration. With the advancement of the "Double Reduction" policy, the education sector urgently needs lightweight and intelligent solutions. Based on this, we conceived the idea of developing an automatic homework grading machine, aiming to alleviate teachers' workload, enhance teaching efficiency, and promote the digital transformation of education through the integration of cutting-edge technologies.Notably, we previously secured a $2,000 GCP sponsorship to accelerate our cloud-based development and model optimization.

What it does

The automatic homework correction machine is an educational correction system that integrates software and hardware. Taking the printer as the hardware entry point, batch scanning of jobs is realized; Relying on Google's Gemini series of large models, the intelligent recognition and correction of handwritten text, chart formulas and mixed Chinese and English content are completed, with an accuracy rate of over 99%. It supports functions such as bullet comment annotation and logical analysis of wrong questions, and generates learning situation analysis reports through the cloud to assist teachers in precise teaching. It has been implemented in K12 schools, colleges and universities, and educational training institutions. A total of over 85,000 test papers have been marked, and the marking efficiency of teachers has increased by more than 60%.

How we built it

Technical architecture

The back end adopts the Spring Boot framework and configures model parameters such as Google Gemini through the AIModelConsts interface. The newapi framework realizes the encrypted storage and dynamic injection of API keys, and supports hot loading and dynamic model switching. The front-end integrates Vue.js to achieve the development of the interactive interface.

Function Development

Multimodal processing:

Utilizing Google Gemini to process text semantics and image information, with the Gemini-2.5-Pro-Preview-05-06 large model enhancing comprehensive understanding ability.

Closed-loop evaluation:

Integrate manual error correction and tagging functions to generate test sets, compare the effects of different prompt word strategies, and automatically calculate accuracy rates, F1 values and other indicators.

Data visualization:

Develop the export function of student situation analysis reports based on the Apache POI library, and generate visual reports such as grade statistics tables, bar charts, and line graphs.

Hardware collaboration:

With printers as the carrier, the entire process of "scanning - cloud correction - printing" is connected, replacing the manual photo-taking and uploading mode.

Challenges we ran into

The recognition of handwritten characters is affected by font style and writing norms, and is prone to misjudgment. The analysis of complex formulas requires the combination of multimodal information, and the technical implementation is highly difficult.

Accomplishments that we're proud of

Technological breakthrough:

Achieving an accuracy rate of over 99% in homework correction, outperforming similar products (such as 98% for Xinghuo Intelligent Marking Machine and 95% for Baidu Zuoyebang).

Market recognition:

It has been piloted and retained by over 10 schools and has been praised by authoritative media such as People's Daily Online and Qianjiang Evening News. The project video has been viewed over 100,000 times on Bilibili and Xiaohongshu, and more than 100 people have voluntarily applied for a trial.

Honors and achievements:

Won the championship of the 2024 AI + Hardware Competition, the Intel Greater Bay Area Innovation Youth Award, and many other industry awards.

What we learned

Technical aspect:

Deeply master core technologies such as multi-modal large model optimization and collaborative development of hardware and software; Accumulate practical experience such as structured output of Function Calling and engineering optimization of prompt words.

Industry perception:

Deepen the understanding of the demands of educational scenarios, and clarify that technological innovation must closely align with actual pain points in order to achieve true value realization.

What's next for Automatic job correction machine

Technical upgrade:

Explore the Intel software and hardware platform and use the Google Gemini series to further optimize the model inference efficiency; Introduce generative AI to enhance personalized correction feedback.

Function expansion:

Develop new modules such as intelligent essay tutoring and oral homework evaluation, covering more subjects and types of homework; Support multi-terminal synchronization (such as mobile phones and tablets).

Market expansion:

Deepen cooperation with educational institutions and schools, and promote the pay-as-you-go model to lower the usage threshold; Plan to expand overseas markets and adapt to multilingual teaching scenarios.

Ecological construction:

Open API interfaces and integrate with other educational tools (such as educational administration management systems) to build a closed loop of the educational digital ecosystem.