Stepan Yakovenko
Портфолио
Inex Technologies
Created production-ready software for NVIDIA Jetson-based ALPR detection hardware from scratch. I have deployed pretrained license plate reading neural networks (YOLO and GRU-based CNN) on new hardware with TensorRT (https://www.inextechnologies.com/wp-content/uploads/2019/11/IZODPU-G.pdf). Implemented inference core in C++11 (video processing as multithreaded OpenCV/ffmpeg application), user web GUI and high-level logic code in nodejs. Also invented optimised CPU-based image downsampler, which significantly overperformed cv::resize in INTER_AREA mode in terms of performance with comparable image quality. Feature-engineered license plate aligner for license plate reader, which performed robustly on skewed and misaligned license plates and allowed to raise recognition precision from ~80% to ~90% on such plates. Took all major technical decisions, managed work of web developer and tester, collaborated with data scientists and technical writer. For another project, which focuses on detecting, measuring and counting axles on vehicles, implemented an algorithm for camera calibration from a single image. Together with object tracking and various 2D geometric approximations, this allowed to measure distance between axles with precision less then one decimeter from a camera, mounted on a toll plaza.
Gazprom-media
Built highload recommender service for online news site with ~300K users per day on average. Used java, mongodb, clickhouse, mapdb, apache hadoop, apache mahout, mlpack. Tried various recommender algorithms and libraries on real users with AB-tests. Used apache sparkml iALS in final version. Solved challenging problem of HTTP responce latency (<20ms per request on average in production under 20rps on peak). Managed to achieve ~20% increase of clicks on recommended compared to top material, ~100% increase compared to baseline random recommendations (measured in A/B tests).
ILFIRON
- Javascript kernel programming for Collab-office, online DOCX collaborative editor based on HTML5 canvas. Chosen technology, created prototype, invented approach for collaborative XML editing, reverse-engineered DOCX format, implemented layouter, solved cross-browser portability issues, implemented unit tests. - Created WebRTC-based web service www.visitele.com with nodejs, which allows customers to embed video chat into their websites and talk directly to customers. Implemented server side in nodejs, client side with AngularJs and pure javascript. Collected requirements from customer, prototyped different approaches, performed highload optimisation for production.