Interview with Ander Steele
Lead Data Scientist, Tonic - Fake data modeled from production.
What is Tonic? Why did you build it?
Tonic.ai is the fake data company. We provide fake data so that developers and QA can test their application code against safe, de-identified versions of their application database.
Why would you use fake or synthetic data instead of real data? Particularly in regulated industries, access to data is very difficult. It may be impossible to test against production data. Even outside of regulated industries, you don’t necessarily want to be working with customer data as a developer. The more distance you can get from real sensitive data as a developer or data scientist, the safer everyone is.
Recently, we’ve been building a product, Djinn, which is focused on the data-science workflow and those use cases. Instead of building an entire de-identified production, de-identified application database, the data-science workflow typically starts with a dataset, which is some view or table within this application database.
The tool takes these views and builds generative models that are capable of producing synthetic data, which has the same statistical properties as the real data. Synthetic data can be used in place of real data to train machine-learning models, build dashboards, to do exploratory data analysis.
This lets the data scientists get started on their job a lot faster than they could if they were waiting around for access to real data. This is something that we’re quite excited about. In addition to the privacy use case, increasingly we’re seeing people that want to use synthetic data for the data augmentation use case.
For example, you’ve paid for some survey data, which has been quite expensive and tedious to gather and that survey data doesn’t necessarily represent the population of interest that you’re trying to study; maybe it skews demographically more females than males. Maybe, the data you have is not necessarily representative of the populations that you’d wish to analyze or the types of questions that you’d like to ask. So one of the things you can do with synthetic data, if you have sufficiently rich and expressive synthetic data models, is then you can generate as many samples as you want – you can modify the distinctions of these various features, and really build alternate versions of your real data that are more representative of the populations of interest.
What does a “day in the life” look like for you?
Tonic.ai is a remote-first company. I’m on the West Coast and we have a lot of people on the East Coast. When I wake up, I usually try to catch up on Slack and see what’s going on - usually, there’s a lot of activity by 7:00 AM Pacific time. There are typically a few meetings each day with my team, but a lot of my time these days is bucketed into R&D — which is trying to improve our models for synthetic data generation — or actually writing features, doing development once the R&D phase is over.
What is the team structure around Tonic?
Tonic.ai employs around eighty people at this point. We’ve grown fairly quickly. I think roughly a third of that is product and engineering, and a significant portion of that is engineers.
Within the engineering teams, the teams are organized around areas of code ownership. Either features or infrastructure. My team, for example, is focused on our synthetic-data product Djinn, which is our tool aimed at data science use cases. We have two data scientists in addition to myself plus a couple of engineers in this team.
How did you first get into software development?
I think it was back in middle school when I started playing around with a variety of things. I think it was Visual Basic, then one summer course in Pascal, which was illuminating.
When I really got excited was on a course in Java during middle school. The most exciting thing about that was writing a little program to render the Mandelbrot Set and of course, zooming in. That intersected nicely with my two interests at the time, which were computers and mathematics.
In high school I got distracted by mathematics, I fell in love with math and pursued an academic career in mathematics for a little while. I still did a lot of coding as part of that to test various mathematical conjectures and numerically verify that these insane ideas actually work.
I was doing a lot of Python programming, mainly because the most relevant software from my area of research was this open-source project called Sage, which is a phenomenal piece of software that can do some really sophisticated number theory computations and many other things. That was my intro to Python.
At some point, I started playing with machine learning. All the excitement in the early 2010s around developments with neural networks and unsupervised learning really caught my interest and attention. I found myself back in the software space doing machine learning. Today, I’m still doing that intersection of mathematics and software.
I don’t use math as much nowadays but It’s fun being able to actually sit down, read these various papers, work out the calculations myself, and try and make sure I really do understand what’s going on.
I think most of my R&D work, aside from figuring out what the ideas are that we should pursue, is really engineering-focused like building prototypes, coding them up, testing them, and doing their whole experimental analysis.
What is the most interesting development challenge you’ve faced working on Tonic?
I’ve spent most of my time over the past couple of years working on Djinn. During that time, the fundamental problem that we faced was how to make realistic, synthetic data.
Let me explain with the help of an example. Let’s suppose that you want to synthesize financial transactions. Imagine you have a large data set of customer-financial transactions. Each of these customers tells a story through time about what they’re doing and when they’re doing it. If you look at it, per customer is extremely high dimensional because not only do you have, for example, the records per observation, but you have all the records per customer. So the dimension of these can quickly explode and the complexity of this problem becomes increasingly difficult with the dimension.
How do you capture those dynamics? How do you build models that are capable of reproducing the statistical properties of these events? We spend a lot of time thinking about this particular problem.
We’ve landed on one of the ideas that’s been floating around in the whole generative AI space for a while, which is using variational autoencoders to make data generators. The setup is fairly simple and abstract. You have your data. What you want to do is you’re going to build some neural network that’s going to take that data, and compress it down into a simpler representation. Take that simpler representation and decompress it back to the data. You do this unsupervised learning process where your network tries to reproduce these input data, passing it through this bottleneck.
You have this encoder and decoder pair – the encoder is taking data, and compressing it into a simpler representation. The decoder is taking that simple representation and decompressing it back to data. At the end of the day, your decoder is a machine for taking noise and producing data, and so that’s your generative model.
What is the most interesting tech you are playing around with at the moment?
It’s a really exciting time right now in the generative AI space. I think everyone has played with one of these very impressive models, either one of the GPT variants or one of the image-generation models like Stable Diffusion.
I’ve been playing around with those models a bit. We built a fun demo using Stable Diffusion for de-identifying images called imafake. In addition to these models, the underlying ideas, if you can call that a technology, are very exciting.
There’s a lot of excellent tooling now that makes it easy to quickly hack. Hugging Face, in particular, provides excellent code that implements many of these research articles and gives you quick access to the cutting edge.
Describe your computer hardware setup
I am pretty deep in the Apple ecosystem, I use the Apple Wireless Keyboard and the Magic Mouse. My display is an HPZ27, which I absolutely love. I’ve recently switched over to one of the newer M1 MacBook Pros, and that’s been impressively fast.
Describe your computer software setup
IDE: VS Code.
Source control: GitHub.
Data Analysis: Jupyter notebooks & PyTorch
Experiment Tracking: Weights & Biases
Describe your desk setup
I have a standing desk, which I sit at all the time, and just some old IKEA chair that has followed me around a few moves, so nothing fancy there.
Daytime or nighttime? Day.
Tea or coffee? Coffee.
Silence or music? Music.
What non-tech activities do you like to do?
The hobby I’m most passionate about is mountain biking. There’s a lot of excellent riding where I live in Santa Cruz, California. Lately, in the interest of using my limited time more efficiently, I’m really excited about trail running. It’s a lot easier to get a good trail run in at an insane hour than it is to get a good mountain bike ride in.
Find out more
Tonic generates realistic fake data for testing purposes. It was featured as an “Interesting Tool” in the Console newsletter on 2 Mar 2023. This interview was conducted on 15 Feb 2023.
Subscribe to the weekly Console newsletter
An email digest of the best tools and beta releases for developers. Every Thursday. See the latest email.