Document Search and ML Training Tool
Product Design Case Study
Note: details omitted to protect company IP.
Note: details omitted to protect company IP.
Problem Statement
A large corporation undergoes hundreds of lawsuits per year and has to sort through, in some cases, millions of digital files to provide legally-required material for each court case. How can we help reduce the number of documents that are sent out to external lawyers for legal review?
Large lawsuits can be extremely lengthy and costly due to the large number of documents requiring legal review and the high cost of legal labor. The corporation wants to reduce the time and cost associated with these large lawsuits.
Approach
1. I worked with the customer to understand their existing process and tools and their major problems and frustrations.
A large corporation undergoes hundreds of lawsuits per year and has to sort through, in some cases, millions of digital files to provide legally-required material for each court case. How can we help reduce the number of documents that are sent out to external lawyers for legal review?
Large lawsuits can be extremely lengthy and costly due to the large number of documents requiring legal review and the high cost of legal labor. The corporation wants to reduce the time and cost associated with these large lawsuits.
Approach
1. I worked with the customer to understand their existing process and tools and their major problems and frustrations.
|
Existing Process
Hypothesis If we train a machine learning model using Relevant and Not Relevant designations, we can predict the relevance of documents and prioritize documents for legal review, resulting in a superior prioritization of documents, reducing the number documents sent for review, as well as potentially catching more documents that would've been missed by using search terms. Proposed Revised Process
|
|
2. I worked with our designer to create Design mock-ups. In our designs, we focused on simplicity, ease-of-use, and interpretability of the machine learning components.
|
|
3. I worked with 7 product engineers to develop the UI, the search backend, and the training system that automatically trains and updates document scores as new information is provided. Simultaneously, I worked with an ML engineer to refine and test the training algorithm on various sample datasets, brainstormed different experiments to run and features to tweak, and reviewed the outputs.
Results
The customer ran a test using the new process and compared it to their old process and found they could reduce the documents needed for review by 50-70% which would save them a comparable amount of money and time. Additionally, they loved the design of the product, citing its simplicity and ease-of-use, which would enable quick and easy onboarding of new users. After seeing a demo of the tool, many other teams within the client organization became interested in using it for their own purposes.
The customer ran a test using the new process and compared it to their old process and found they could reduce the documents needed for review by 50-70% which would save them a comparable amount of money and time. Additionally, they loved the design of the product, citing its simplicity and ease-of-use, which would enable quick and easy onboarding of new users. After seeing a demo of the tool, many other teams within the client organization became interested in using it for their own purposes.
|
|
Though our first customers were in the legal space and the tool was initially intended for use in the legal space, we knew there were many other potential applications for this product. Many industries are presented with the problem of having a set of documents in which they want to find important information. We have since expanded to work with other customers in different industries to solve this same problem.