AI image compression
to decrease image storage and transfer costs with fast decompression on edge devices
to build brand-new entropy reduction transforms, we used highly optimized AI predictors and post-processing filters
the technology we developed was tested for on-chain image storage. The cost of storing images directly on the Ethereum chain was reduced by up to 100 times compared to PNG
stack: Adaptive quantizers, entropy reduction, lz77, predictive coding, entropy coders, Pruning, QuAntization, Apple Neural Engine
huawei smartphone ai camera
to develop an AI HDR solution that matches the quality of the baseline model and can process a single 4K photo in less than one second on a Qualcomm 888 NPU
we built a new mobile-efficient 8-bit architecture that accounts for the physical noise model on the camera matrix
deployed AI HDR cameras’ software in Huawei P50 and P60 smartphones sold in hundreds of millions devices globally with a preset photo application
stack: Read&Shot noise models, homogeneous functions, QuAntization, Pytorch
* Wallarm
is one of the leaders in security products for cloud-native environments. Graduated from Y-Combinator's S2016 batch, this company competes with its AI-based product against the likes of WIZ, ORCA, and others
wallarm* ai firewall
to create a solution that automatically reconstructs web application endpoints and classifies input requests as safe or malware in real-time, achieving a bandwidth of 1 Gb/s on a single CPU core
we developed a mobile-efficient 8-bit architecture that accounts for the physical noise model on the camera matrix
we deployed and transferred a C++ library, which includes ML tools we developed from scratch, for HTTPS malware detection. This significantly improved both accuracy and speed
stack: Categorical stat tests, LOCK-FREE MAPs, tree data structures, REGEX LEARINING, C++, CYTHON, DOCKER
Recognition of brain signals for Nissan cars
to create a solution for Nissan cars that detects what the driver is interested in and determines where their attention is. Can we deduce this from the driver's brain?
we recorded brain reactions
to different traffic signs during simulated driving and classified them as Target vs. Non-Target
compared to previous methods, we improved the accuracy of street sign recognition from brain signals by almost 7%
stack: C++ (driving simulator, eye-tracking) MATLAB (EEG, eye-tracking), Python (EEG)
