Start-up in robust and safe AI, funded by Business Finland with €0.7M.

Worked on evaluating the susceptibility of AI solutions to adversarial attacks.

Adversarial machine learning and AI. Improving AutoGAN defense against adversarial attacks, including the CIFAR10 dataset and more types of attacks in the experiments. Took part in the successful first phase of DARPA’s AlphaDogfight Trials' competition with simulated combat between F-16 aircraft.

Working on adversarial machine learning, I developed a novel way to counter adversarial attacks that is able to handle different kinds and magnitudes of attack. In addition, it defends effectively even in the lack of adversarial training points. Accuracy on MNIST dataset exceeds FGSM method for different magnitudes of attack for training and testing.

Improved classification results of various datasets by utilizing mutual information between features. We generated new features as regressions of each existing feature from the rest of the features and then achieved better classification results from all features combined. The top improvement on the error was over 27% points.

Working with Nvidia's autonomous driving team, I applied regularization to the deep learning network resulting in a decrease of 91% in the size of the network performance, accuracy improvement of 8% points and reduced over-fitting. The decrease in network size that I implemented was critical to the project because it decreased the memory required for deploying the autonomous driving solution. I also led efforts for a crucial component of the learning system towards its reliability, innovating two metrics for establishing a reliability estimate for the output of the model.

Enhanced functionality of an emergency notification system: I added voice input to the emergency location, used text-to-voice tools for customized alerts, made the interface more intuitive, implemented an internal messaging system for better response time. My work was part of an effort that led to a patent for an emergency alert system.

Created an on-line tool for analyzing activity of cell phone users. Implemented interface and functionality requirements to a social networking website. Responsible for back end data processing. Tools: Linux OS, PHP, Apache, MySQL, XML, Anychart, JavaScript, CSS, HTML.

Implemented TCP-option negotiation for a cryptographic authentication protocol, as well as support for distributed lookup services. Extended functionality of stateful packet filtering firewall. Tools: Linux OS, C, networking protocols (TCP/IP, etc.), VMWare, Svn, Awk, multi threading.

Designed and implemented GUI for easier management of WiFi connectivity for Nokia 770 tablet. Improved the existing GUI for WiFi connectivity in the Nokia 770 tablet. Tools: Linux OS, Maemo environment, C, MySQL.

Gathered and analyzed requirements, designed and implemented bank inventory software. Designed and implemented software for handling of treasury bills. Database design and implementation. Tools: Windows OS, Visual Basic, MS SQL, Crystal Reports.

Extended functionality of software for generation of customs documentation for shipping goods. Tools: Windows OS, C#, .NET, MS SQL, List & Label.

Solely responsible for municipality software: gathering and analyzing requirements, design, implementation, deployment. Extended functionality of accounting software. Tools: Windows OS, Visual Basic, MS SQL, MS Access, Crystal Reports.