IT Arena is a place where the most extraordinary tech meetings happen – a dose of inspiration you get after the event is unparalleled. In this blog, we’d like to share highlights of the 5 tech talks that took place at the 2019 edition of IT Arena.

Pavel Zak – Building a game in SwiftUI – A Madness? 

Pavel Zak has a vast experience in software development. He developed several mobile apps for smart homes, games, traveling, etc. Previously, Pavel worked as Chief of Mobile Product at Kiwi.com. Currently, he is working on his own app in SwiftUI – a new toolkit for building user interfaces for iOS.

With SwiftUI you can compose and change view modifiers, layout position, masking, background, foreground, appearance, blend modes, animations, and transitions. The most powerful feature of SwiftUI is an animation system. You don’t need to write complicated animation code blocks.

“I want to build a bit of hype about SwiftUI – a great tool to write applications and board games. It simplifies your app architecture by providing environment objects, accessible for every subview. A library with 1500 flexible icons, called SF symbols, which change the size, depending on user preferences is at your disposal,” says Pavel.

Ian Patrick Hughes – Hacking Humans with A/B Testing

“Facebook, Google, Instagram, and YouTube are already hacking us, even this morning, I’m sure,” says Ian.

To capture human behavior developers need to know three things about the user: motivation, action, and trigger. After this, developers can try to design good A/B test with the following six steps:

• collect data about users;
• identify goals;
• form a hypothesis;
• create variations;
• run the experiment;
• analyze the results;

There are various A/B testing solutions: Google Optimize, Optimizely, AB Tasty, SiteSpect, Adobe Target, VWO, Sentient Ascend. Most of them are client-side scripts, which means you have to inject JavaScript code into your existing HTML pages. With little snippet you include, users will get a randomized page so you can confirm the hypothesis and select which solution works best for you.

Veronika Demedetska – Robot Simulation from Scratch

“People nowadays use more than three million industrial robots. Not all robot manufacturers have their own simulation software, and even if they have it, they can’t cover all customer needs. As long as robots can’t survive without software developed by humans, we still have a job,” says Veronika.

People use industrial robots for heavy physical labor like welding, painting, assembly, packing, labeling. For the implementation of a simulation system, it’s necessary to do several steps:

• 3D modeling and rendering;
• movement simulation: robot controller, forward and inverse kinematics;
• collision detection algorithms;
• learn how to teach the robot;
• set and save robot position;
• apply base generation algorithms;
• make 3D vision systems that simulate different kinds of operations.

In the end, it’s possible to simulate all production lines, the robot itself, or the environment. It’s crucial to describe the whole scene for creating a simulation system: the robot and surrounding devices by TCP – tool center point that helps to navigate.

Mohamed-Achref Maiza: AI for car diagnostic with TensorFlow and Google Cloud Platform

As a Data Scientist at Renault Digital, Mohamed-Achref Maiza is responsible for two kinds of deep learning applications. The first is about automating quality control in manufacturing with deep learning, and the second is about automating car diagnostic, using AI tools and Google Cloud Platform.

“Imagine, you have all data about the vehicle – metadata, car model, gearbox type. And you also have specific data about the diagnostic – error codes, symptoms, context values. Given that, we want to create a system that can predict operations. If you predict that the technician should replace the battery in the electric car, but it shouldn’t be replaced, we are paying not only the cost of the battery but also the cost of the new investigation,” illustrates Mohamed.

It’s good to use TensorFlow – open-source library for developing and training ML models. The API makes it easier to read data from many sources like TFRecord files; this is like a specific data storage format for TensorFlow. For the production environment, the excellent option is to train models, and monitor workloads on the Google Cloud Platform. To integrate ML into your pipeline, use Kubeflow – the platform dedicated to making deployments of machine learning workflows on Kubernetes simple, portable and scalable.

Alex Peattie – Using NLP to uncover the hidden treasures in your data

Something fascinating is happening in the world of NLP. The large visual database for software research ImageNet and transfer learning moment have now arrived. People have a massive pool of unstructured data that is not used. However, humans can take that data and turn it into a structured one, get new insides, and make our organization better.

“A few tips if you want to get started with NLP. Fast.ai has a bunch of good tutorials. TensorFlow has built-in support to download models in the line of code. HuggingFace provides many models ready to download and be used,” recommends Alex.

You need to train transfer learning models on the language data. You can download the model and fine-tune it, whatever your particular task is. It started with ULMFiT and then moved to a couple of bigger models like ELMo, OpenAI, GPT-2, or Bert from Google that trained on about three billion worlds from Wikipedia.

More speaker slides from IT Arena 2019 can be found here.