Part III: Ask a Machine Learning Engineer with Sensibill’s Antonio Sou
In Parts I and II of our Machine Learning series, we explored how the team here at Sensibill applies ML and deep learning to read and understand important receipt data. (Oh, and we dove into how it all works, too!). If you love Machine Learning or want to work at a Machine Learning startup, this series is definitely for you!
In this article, we’re doing things a little differently. We’re giving you a glimpse inside the mind of a Machine Learning Engineer, diving deeper into the topic and what it means to real Sensibill team members like Antonio Sou - one of our amazing Machine Learning Engineers.
Let’s take a look at where Antonio feels Machine Learning can benefit businesses and why you should consider joining one of our Machine Learning teams here at Sensibill!
Why should financial institutions care about the benefits of Machine Learning when it comes to their digital product offerings or services?
Machine Learning can be of significant benefit to FIs because it can automatically analyze the vast amounts of data that they process every day. For example, if a financial institution could harness AI to reduce fraud even just by 1% or 2%, that can have a huge impact for a bank. That may not seem that substantial, but AI is continuing to drive value for banks; some reports are even showing AI technologies could deliver up to $1 trillion in added value to banks every year.
As a Machine Learning developer, there’s a great opportunity to make an impact using ML and AI, not only for financial institutions but also for their end users, helping them build and foster financial wellness in their everyday lives.
What do you think is the greatest challenge for any Machine Learning expert in an industry like fintech?
I think it's building the trust to get access to the users' data. It's super-sensitive data. And so you have to have processes in place that protect that data, and you have to be able to demonstrate that you can do something meaningful with the data in the first place.
In our case here at Sensibill, we have multiple training sets and validation sets for our models, but we also need to propagate specific governance requirements from financial institutions. Building that trust and working with those requirements definitely adds a level of detail around data sets that rarely exists in other industries. This means we’re treating our clients’ data with the same level of importance and trust as they do.
What do you think is the most valuable data when it comes to our industry?
This depends on what you're trying to accomplish. I think you can do a lot with item-level transaction data and browsing history (what kind of items a user was looking at, for example).
Historically, it's been challenging to really understand what an item means, and we've relied on keywords to derive meaning. But keywords have their limitations because items appear differently across merchants; some use SHRTND versions of words, or sometimes we have OCR errors (to name a few).
With the recent release of pre-trained BERT models, we can now create resilient word/ sentence embeddings that take into context the entire knowledge of the internet! This may not sound like a big deal, but it's incredibly powerful and allows smaller tech companies that don't have Google's scale to do really cutting-edge NLP.
Not to mention, the capabilities we have with ML and AI mean we can enable both large-scale and small to medium-sized organizations to help their customers move towards better financial wellness (like decision making, or reporting on their purchases, to name just a few). Tie in the personalization of services through data, and ML and AI are key technologies that can really benefit our clients.
We're just scratching the surface - if our work and solutions speak to you, come dive deep with us into the world of Machine Learning in fintech.
By the way, we just released our newest digital solution: Receipt Extraction API. You can also check out our article to learn more about why we created this API and how our machine learning powered-technology can help businesses extract better data (right down to the SKU-level).
Where do you think the fintech/financial services industry needs more education around Machine Learning?
Machine Learning is not as scary as it used to be 5 or 10 years ago. There are lots of platforms out there that let you get up and running really fast. I think the focus has now shifted a little bit from hard-core academics to hard-core engineering. When building ML teams, it's paramount that they have a background supported by both strong coding/engineering skills along with research.
Do you see any emerging use cases for Machine Learning or near trends in ML?
There's definitely a trend of automation using Machine Learning. Every year, AI can do more and more things better than humans. The next big shift that will affect society the most (in my opinion) is the automation of transport (like self-driving vehicles).
How or why did you become interested in Machine Learning?
It's an incredible tool that allows us to build incredible things. Similar to how regular programming opened lots of doors 20 years ago, AI is opening lots of doors right now.
What's your favourite algorithm?
Well, I'm an engineer, so it's whatever algorithm is best at solving the problem! But I think transformers are really interesting in deep learning because they unlock a bunch of use cases around NLP (think BERT or GPT3), and it's easy to leverage pre-trained models using transfer learning.
What's more important for you: model accuracy or model speed?
Again, I'm an engineer, so it really depends on the application and the problem we're trying to solve, though I'd say both are important. A Machine Learning Researcher, for example, cares about state-of-the-art accuracy, and would use those models for background tasks. For a Machine Learning Engineer, we’re focused more on productization, making speed important for delivering the best user experience while scaling millions of users.
Why should I join one of the Machine Learning teams at Sensibill?
There are some hard challenges to be solved and a variety of things you can work on. Take the example of extracting information from an image of a receipt. There are many steps that need to be working together—the quality of the picture the user takes and ways to enhance the image, which directly impacts the OCR quality of the image, which directly impacts how good your NLP and extraction can be.
There are many ways to optimize this pipeline, including human factors, computer vision, OCR, NLP, and sometimes, just smart algorithms that can enhance each component. ML and AI drive Product at Sensibill in ways that I haven’t seen at other companies. For example, we actually use our model error rates to drive product decisions, and vice versa. ML is a core partner of product, which is rare among tech companies.
Interested in becoming a member of one of our Machine Learning teams at Sensibill? We’d love to hear from you! Click the button below to explore our career opportunities.