Machine learning TL у Google

ПРО МЕНТОРА

Приблизно десять років я розробляю платформу для машинного навчання в компанії Google, а також застосовую цю платформу для покращення алгоритмів вебпошуку, YouTube, Play Store та інших продуктів. З-поміж іншого, розробив модель для оцінки якості застосунків в Play Store, модель для оцінки якості відео на YouTube, модель для передбачення часу перегляду відео на YouTube, та багато інших проєктів.

Цікавлюсь перенесенням ідей, практик і процесів з розробки програм у машинне навчання. В минулому — любитель змагань із програмування й адмін на TopCoder.

ЕКСПЕРТИЗА

Machine Learning
Python
C++

ДОСВІД РОБОТИ

Google
Staff Software Engineer
May 2020 — Present (2 years 2 months)
Building the best Google Assistant.

People.ai
Staff Software Engineer
April 2019 — May 2020 (1 year 2 months)
San Francisco Bay Area

Google
Staff Software Engineer
January 2012 — April 2019 (7 years 4 months)
Mountain View, CA
* ML engineer and TL in Ranklab: a MATLAB-like data science platform environment which is behind ML models for most Google products.
* Creator of new tools and approaches to debug, understand, and improve machine learning models.
Led design and improvements for ML systems behind several dozen of biggest Google products.
* Led the team to redesign the model used to predict the quality of all apps in Play Store. Increased model precision by 10% at the same recall rate. The model is used across all Play Store properties, and was an important part of the yearly Play Store OKR.
* Improved the model predicting quality of videos for all YouTube search queries, which contributed to more than half of the yearly OKR.
* Launched significant changes to ranking formulas for other Google properties, including Youtube WatchNext recommendations and AutoPlay, GMail, Drive, parts of Web Search, and many smaller ones.
* Grown the team and led it in multiple ML projects.
As a software engineer on the same platform:
* Led the efforts of integrating the platform with TensorFlow ecosystem.
* Designed and implemented storage data format, used to store all training data for the platform. Later, improved the format halving data footprint and loading time.
* Integrated the platform with the system for model evaluation. My efforts allowed analysts to load any slice of data to a coLaboratory notebook for detailed analysis.
* Redesigned the system for debugging ML models, allowing engineers and analysts to interactively analyze models and compare them against each other.
* Oversaw development of the user-visible APIs, and documentation for the platform.
Site Reliability Engineer
May 2010 — December 2011 (1 year 8 months)
Dublin, Ireland
Part of the team responsible for keeping distributed storage systems (Google File System, BigTable and Colossus) running.
Significant accomplishments:
* Automated cluster management process.
* Proposed, designed and implemented an automated approach for storing cluster configuration.
* Improved resources provisioning for storage software company-wide.

LunaDev
Independent Flex developer
December 2009 — June 2010 (7 months)
Developing games in Flex to entertain myself.

Elyland, Ltd
Java/Flex Software Engineer
March 2009 — April 2010 (1 year 2 months)
Kyiv, Ukraine
Development of a MMO strategy game (10k players total / 2k players online handled by 1 machine). The project successfully launched, quickly becoming the most popular online game in Ukraine and Russia (150k players one year after launch) and winning several awards.
Technologies: Java (server side) and Adobe Flex (client side).
Supervised the project from algorithmical point of view.
Significant accomplishments:
* Designed and implemented architecture to handle all actions involving more than one player.
* Debugged through and resolved several major performance bugs.
* Generated multiple ideas on improving SCRUM process.

TopCoder
Algorithm Coordinator
August 2005 — February 2009 (3 years 7 months)
Coordinating the process of preparing the problems for multiple algorithm competitions. This includes: writing and testing problems for competitions, managing problem writers and testers, organizing the workflow, etc.

Flextronics
Java Software Developer
February 2005 — July 2005 (6 months)
Several small projects.

Materialise NV
Software engineer
August 2001 - December 2004 (3 years 5 months)
Working on an Rapid Shell Modeling project, which digitized creation of personalized hearing aids.
Significant accomplishments:
* Designed an implemented API to handle manual and automated positioning of 3-D objects relative to other 3-d objects.
* Developed an algorithm for optimal positioning of complex 3-D objects.