Accelerating document classification with Machine Learning

Berry Recruitment engaged STX Next to develop a Proof of Concept project that leverages Machine Learning to efficiently classify documents shared by job seekers. To optimize user flow and ensure smooth integration for both end-users and internal stakeholders, we teamed up with Product Designers.

Addressing the challenges

Berry Recruitment partnered with STX Next wanting to build an AI model that could classify uploaded documents into different categories.

These documents were usually scans or photos of physical papers, complicating the classification task.

An added difficulty was the lack of standardized labelling in the company’s existing document database, necessitating additional efforts to achieve proper labelling.

Building a prototype

To solve the document classification challenge for Berry Recruitment, we implemented a two-step solution:

Labelling Process

The primary step was to identify distinct classes based on the file names. We categorized approximately 46,000 documents using various labelling strategies, providing a comprehensive dataset for training the image classification model.

Image Classification

For the classification part, we employed a pre-trained Convolutional Neural Network model (EfficientNet) for image feature extraction. Then we added two fully connected layers for the final classification.

—

Metrics and models used: Accuracy, Macro Average F1-score, Most Frequent Class Voter, Uniform Voter, and Stratified Voter.

Want to turn your idea into a real product?

Reach out to our experts and let us guide you through the process.

Get started

From PoC to the final solution

The insights from this PoC project are set to influence the development of a full production solution, ensuring scalability and greater efficiency in the long term. Our client gained a reliable pathway to improve their document handling processes with:

Enhanced Accuracy

The most effective model iteration achieved an Accuracy of 91% and a Macro Average F1-score of 89%, outperforming all baseline models. This indicates that the neural network learned effectively, proving its potential for production deployment.

Operational Efficiency

The PoC provided concrete evidence that incorporating AI techniques could substantially enhance the document validation and verification process. The compelling results showcased that automating this task is both feasible and beneficial.

User-Centric Design

In addition to the Machine Learning aspects, we delivered customized user flow designs aligned with the AI model's approach, providing the most convenient experience for end-users.