Working on a solution (Java) that allows users to train / deploy ML models for data extraction (models are python / torch based, delivered as Docker images, managed via a Kubernetes backend).Lots of troubleshooting leading to fixes and improvements to the solution.Some notable activities: discovered (and fixed) improper use of green threads in pyhton code leading to abnormally large request duration under high load, implemented a tool-set to profile models in prod-like environments, decisive contributions to fixing CVEs, discovered a race-condition in torch-vision that prevented concurrent running of predicts, improved model speed by page-level parallelism, improved GPU utilization (and ultimately speed) by fixing data pipeline bottlenecks, found and fixed source of high memory consumption in data pipeline.As POCs: re-architected model to achieve maximum GPU utilization, re-architected model to allow serving multiple user models on the same system (loading up to 20 user models per GPU, also allowing hot-swapping).Special focus was on the data pipeline optimization via re-factoring and by leveraging multi-processing, leading to improved GPU utilization both at inference and at training time.

Driven the backend effort of porting of the Adaptive Analytis Manger to a Web based UI.Implemented feature correlation test functionality using Cramér's V in Adaptive Analytics Manager (a PRM plugin) as a tool for the modelling process. The implementation is mostly SQL code (plus a nice heatmap in the Java UI)Contributed to porting the proprietary scoring engine to an Openshift environment - supporting Kubernetes probes, full configurability, metrics etc. Implemented a new native REST interface using Spring Framework for a new product integration.Implemented a "distributed feature repository" functionality into PRM (a fraud prevention product for financial institutions) using REST services with Spring Framework.Took ownership of a proprietary scoring engine, being in charge of releases, fixes, enhancements and integration issues with the main product.Implemented the changes needed to interface the main product with a third party scoring engine (custom requirement from a specific client, but shipped as a feature of the product).Involvement in different architectural changes to streamline the machine learning flows (both production and data gathering for model building).Keeping in touch on a regular basis with the data science team to learn their specific processes and requirements, so the software engineering team from Timisoara has better insight when making architectural changes needed to support the machine learning modules.Contributed to the research for the incremental learning initiative by implementing and benchmarking the AVM algorithm (Approximation Vector Machine) in collaboration with the Data Science team. Contributing to development of a reason processor in collaboration with the Data Science team.

During my 10+ years at Bitdefender I worked on several projects using different technologies and frameworks/environments, as an engineer and as a team lead. Highlights:- research the use of an external device to host the scan engines, so the main computer CPU is offloaded. This task involved a lot of C/C++ on linux and Windows platforms, since the box run linux and the integration was made into the existing Windows product- create a distributed and scalable system that scans emails as a service. The scan engines were integrated into a "processor" machine, with POSTFIX frontends. The data was stored in a mongodb database (also RabbitMQ)- research the capabilities of vShield Edge and vShield Zones and how they would be a match for being configured by the Bitdefender management console (lot's of web-service calls with some private APIs, not very well documented)- implement a system to manage iOS devices from Bitdefender's management console (for the Business line of products). Research Apple's MDM (Mobile Device Management) protocol and system, implement a management system that meets the requirements of the Business product. Technologies: node.js and mongodb.- implement an application for Android devices that offer management of the devices in the Business environment: it integrated Bitdefender's scan engines for Android, it could deploy WiFi credentials, force the use of a passcode, locate / lock / wipe a lost device, monitor installed applications, control web-access etc. based on security policies applied from the management console.- implement new features in the Endpoint Security Client for Mac for the business line. This involved extending the existing integration API for third party integrators, fixing existing bugs and features like Silent Mode (no UI visible to the user), a scan task scheduler, a module that can apply different settings based on the current user and network, new modular architecture for application etc.

Worked as a Team Leader for a software development team. The team developed several new features for the Console (UI) of the InfoPrint 4100 Printers.

Started as Junior Software Developer and worked my way up to Team Leader. Tasks included development of web applications using CGI (C/C++), developing utilities for internal use in Visual C++, development of a graphical editor for web applications in Java (two versions) for a proprietary eCommerce application, development of desktop software in Visual C# and development of a web site using ASP.NET (including interface to a proprietary database, credit card processing etc.)