The Problem Space.
Traditionally, companies hire teams of data scientists, machine-learning engineers, and SMEs to build complex machine-learning algorithms to automate document-heavy workflows. However, data scientists are scarce, data and business requirements are constantly evolving, and building highly accurate, reusable automation solutions is time-consuming and costly.
Why not speed up the time-to-market by automating as much of the model training process as possible, and put this technology directly into the hands of the average business user?
Challenge
Design from scratch an intuitive UI that enables users to build, train, and fine-tune machine-learning models that can classify and extract key data from structured, semi-structured, and unstructured documents.
My Role
UX Designer
My Responsibilities
Identify and prioritize key user type(s)
Design user interface that enables user to efficiently train a machine-learning model
Conduct usability testing to validate product design
Gather user feedback to continuously iterate upon design and identify UX improvements
Results
Platform used to train models for over ten use cases
Platform generated ~$300k in enterprise, partner, and reseller deals
Who are we designing for?
To identify the key user types for our product, I partnered with our business team to identify our core customers, which include Global System Integrators, IT Consultancies, BPOs, and automation departments within a corporation. Each customer has different motivations— some, to integrate our solution (backend) into their existing platform, and others, to use our platform to build machine-learning models to tackle their use cases. Many of these businesses already have existing automation practices but are looking for better solutions.
I looked into each customer type to identify the actual users behind the automation practices in place, and identified three key user types.
User Flow
Since I was new to the AI space, I worked with our product manager to understand the steps the user needs to take to train a machine-learning model, and mapped out what the primary user flow would look like
My Process
After identifying our user types and the user flow, I followed the process below. I broke down each primary goal (e.g. training a model) into smaller goals, always keeping the user’s wants/needs in mind (What would the user expect to see on this page? How can I make these features clear and discoverable? How can we automate as much of the model training process as possible?)
The Solution
UPLOAD & CLASSIFY
Users have the ability to upload, classify, and manage their files on the Files page.
DATA LABELLING
Users can quickly label their data using the labelling tool on the side.
In the first iteration, after clicking “submit”, a green check mark would appear next to the file name. This resulted in a long list of “submitted” and “not yet submitted” files which took more time for the user to sort through. To increase labelling efficiency, I created the “Submitted” and “Unsubmitted” categories which organized the files accordingly so the user could complete their labelling goal quicker.
MODEL TRAINING
Users can easily train and fine-tune their AI models on the model training page.
Users can create new models, fine-tune existing ones, and select which model iteration to set as the active version.
The model loaded on this page can be applied on the labelling page to assist the user in the labelling process, speeding up the overall model training pipeline.
