TNG goes virtual
Sometimes, we have to make a virtue of necessity. In view of the spreading SARS-CoV-2 virus TNG started an experiment last Friday, a “Virtual Techday.” This consisted of virtual workshops, presentations and even a virtual “Weekly half-hour,” an informative in-house event hosted by the TNG partners. The Weekly Half-hour had previously been streamed, in addition to the actual presentation. However this was the first time as a purely virtual video conference. Up to 230 colleagues attended simultaneously, applauded via emoticons, and posted comments in the chat.
The feedback on the virtual Techday experiment was extremely positive. However, we are all looking forward to meeting again in person.
Remedy for Performance Issues
Black Friday: every year this top-selling day poses a unique challenge to our rapidly growing customer in the online retail industry. The webshop must operate reliably under an unprecedented load. Despite a hardware upgrade, in 2019 there were already short service disruptions observed at peak times prior to Black Friday. Trouble was clearly imminent.
We identified and resolved multiple performance bottlenecks. The configuration of the application was then optimised in an iterative fashion. This was done using a JMeter test plan with request distribution modelled on customer behaviour. Executing it from the cloud directly against the production servers enabled us to test different load scenarios. At the same time, technical and organisational measures were taken to ensure that customer shopping remained undisturbed. Elastic Stack, Prometheus & Grafana served as the main tools to identify and analyse backend metrics.
Ultimately the improvements yielded the desired result. The throughput doubled, but the system withstood this crucial phase without any issues. The new monitoring processes allows future performance problems to be detected, understood, and rapidly solved.
Deep Fakes in real time
Deep Fakes are imitations of pictures and video, created using artificial intelligence. Common scenarios include exchanging faces in order to create the illusion to see another person in a video. Within the scope of a research project, the TNG Innovation Hacking team investigated Deep Fakes to better understand what is technically feasible. Furthermore, we wanted to explore the limitations of the technology - especially if and with what image quality it is possible to create Deep Fakes in real time.
Our research shows that it is possibled to exchange the face of a person filmed via a real-time video streamwith the face of another person, including applying facial expressions and movements of the person filmed.
By applying various techniques from the area of computer vision and neural networks, faces in the video feed are recognized, transformed and embedded in the video output. The project uses autoencoder networks trained in Keras. They were trained using so called GANs (Generative Adversarial Networks). Aditionally, the developers used different neuronal networks for face recognition and segmentation.
AI Size Recommender for Fashion
For a large fashion retailer, TNG developed an application that makes individual size recommendations for web shop customers. The task for the solution was – despite a large number of size ranges and partly low availability – to display to individual customers only those products really likely to fit. To do this, we started with an idea workshop, convinced our client by developing a prototype, and finally rolled out a production solution with measurable success. To create the individual size profiles, we experimented with various machine learning approaches. The final version was then implemented using AWS Lambda and Elasticsearch.