Amalgam

Designing a machine learning platform to enable coworkers from different levels of expertise train their own data in an easy flow.

Company

Exponential Ventures

Year

2020

Role

As a Product Designer, I was charged with the design of Amalgam platform from the ground up. I coordinated and led all the steps of the design including: interviews, user flows, interaction, visual, product, prototyping and participatory design in order to ensure the result meets company's needs and is usable.

Overview

Exponential Ventures felt the need of having their own machine learning platform that could be used by members of medium and high level of expertise in the field. The result of this project is Amalgam, a machine learning platform with a seamless flow that trains prediction models. To develop it was used design thinking applied on product design following the five key phases: Product definition, Research, Analysis, Design and Validation. To ensure that the product would meet company's need, we added participatory design in the methodology. After constant back and forth collecting feedback about user flows from software engineers and stakeholders during the Design and Validation phases we came up with the prototype and after several tests, the implementation of Amalgam. Amalgam is now live and available not only for Exponential Ventures but for other companies that would like train their data.

The Challenge

This was my first project as a junior product designer, the challenge starts there. Even though I use AI as a tool everyday, I did not know exactly how machine learning platforms work. Since the company works with artificial intelligence, quantum computing and digital fabrication, the need for training data increased. The team felt the need of a platform that would easily train prediction models and help improve company's products.

Goal

A platform that would attend and both users with medium and high level of understanding of data training.

Constraint

The main constraint was that the number of users to interview and validate ideas was limited to people that work inside the company.

The Solution

The method used was design thinking applied on product design. This process helped me validate every step of the process even though I had only a few users to interview. With that in mind, the steps was divided by Product definition, Research, Analysis, Design and Validation.

Design thinking applied in product design

Product Definition

Interviewing CEO and stakeholders to gather insights about business goals. They felt the need of a machine learning platform in order to deliver faster products to the market. Moreover would be more financially beneficial have its own platform instead of relying in other companies. The goal was to have a platform that would attend both users with medium and high level of understanding of data training.

Requirements

The platform would offer 3 types of models: Classification, Regression and Clustering.

Users

Data Scientists,

Software engineers

and CEO of the company.

Research

Competitive research to understand industry standards and identify opportunities for the product we are building. Our direct competitor was Amazon SageMaker. Our (not-direct) competitors was Microsoft Azure and Google ML. Due to the fact that is needed to pay to access those platform, I relied on screenshots and videos from competitors to see the design and understand how their platform works.

The flow of competitors platforms

Analysis

It was created two personas that would represent different levels of expertise in machine learning. One is an expert and other one a mid-level experienced person.

Personas with different level of expertise in machine learning

Design & Validation

I included together the process of design and validation in this use case, because this project was a constant back and forth until we refined to a well done product. Most part of the features of the platform were something that I had never seen been designed before. One example was the "stripping" feature of Amalgam, where is used to clean data. I learned that I did not need reinvent the wheel. Even though some features are hard to design, you can always check examples of patterns that already work and start from there.

Features and flow

While designing Amalgam, the amazing development team was already building features that would be included in the platform. Most of those features would not be visible to the user, they will run automatically in the platform.

Elemental: Analyse the data that is being processed and verify if there are conflicts.

Stripping: Points any conflicts and help the user to clean the data before running any project.

Aurum: Keeps track of all user experiments and its performance and gives the user ability to export any experiment straight out terminal or github.

Spooky: Allows the user, when provisioning one experiment, to stop or restart the provision as long as he wants.

Based on those features and analysing the steps needed to train data, we built this flow. Importing and cleaning data can work independently of a project running or not but a project can only run with a dataset that is clean and ready for a project.

Amalgam plartform flow

Prototype

Even with low fidelity wireframes - which was not so low fidelity because I created inside a digital platform instead of going paper first - we would test the flow and progress every Friday. So every wireframe that I had I made it clickable in order to increase the fidelity of the flow we wanted to achieve. At first we added too much options to upload the dataset in the system, we simplified that by allowing just a few ways for our MVP. The cleaning data phase also had a lot of versions because every Friday with would fix a problem and realising we were opening another one when working with that. After experiencing different problems, we evolved our product and delivered a good flow.

Amalgam interactive prototype

Results

Amalgam is already available and you can ask a demo on amalgam.ai. During this process I learned how to deal with features that I had never saw being designed before. Building this product was a great way to kick-off my product design career. I starting by not knowing exactly how a machine learning platform works and I ended up with the understanding of the flow around all data training.

The next steps is to transform the platform in a self service machine learning platform. This will definitely require interviews with users, we cannot rely on our own anymore. This is just the start and I believe the format of amalgam will change a lot throughout the years.

This is one project that I am very proud of. Amalgam thought me to face challenges as they come. Moreover it was a hell of experience to create a product where I had so limited users to interview. Looking now what I would do differently is:

Test with people outside of our company, since building a product and test inside the company can create a lot of bias and end up just fixing problems that are inside our bubble.

Sketching way more wireframes on paper in order to save time instead of doing directly on the screen. It's inevitable, once a designer go to the screen we feel the urge to add buttons, texts and more details. I could have saved more time avoiding that and not worrying with those details in such an early stage.