Anton Dergunov

Senior Applied Scientist @Microsoft

London, GB
MOBILE NUMBERS
+91 *********19

Signup · Get unlimited contacts

WORK HISTORY

May 2018 — Present

Senior Applied Scientist @Microsoft

View department →

London, GB

Applying machine learning in Microsoft Office products:* Enhanced the quality semantic (natural language) search for people/contacts in Copilot. Performed evaluation of embedding models and cross encoders, leveraging LLM-based assessments and synthetic data generation.(LLMs, ChatGPT, RAG)* Built from scratch the pipeline for collecting \"eyes-off\" training data, training ranking and classification models with offline evaluation. Optimized the pipeline\'s performance to process large volumes of click log data. The model is used for people ranking in many Microsoft products. Deployed several models to production, demonstrating online metrics improvements (such as up to a 4% increase in click-through rate; improvements in time to success).(Azure ML, Python, PySpark, LightGBM)* Evaluated the Personalized PageRank algorithm over organization people graphs for this people ranker, showing a 2.15% increase in NDCG@1 and a 1.21% increase in NCDG@3. Tested other graph-based approaches produced by different teams.* Enhanced the relevance of recommendations in the Meeting Insights feature of Microsoft Outlook. This feature recommends emails and documents relevant to meetings.(Python, SciKit-Learn, NumPy, Pandas, PyTorch, LightGBM)* Enhanced the data extraction pipeline, increasing the amount of training data by ~50x, which improved model performance (AUC up by 3%, AUPRC up by 12%). Further data analysis improved model performance (AUC up by 3%, AUPRC up by 4%, NDCG@3 up by 11%) and enabled twice as many features.* Evaluated new features for this machine learning model by identifying and optimizing new signals (e.g, social, content similarity, key phrases).* Performed experiments using the BERT model and related NN architectures to improve content similarity features in Meeting Insights: verified the model on multiple datasets, performed error analysis, comparison with existing features. Distilled BERT into simpler models. The model showed a 3% increase in NDCG@3.

EDUCATION

2008 — 2011

State University of Nizhni Novgorod named after N.I. Lobachevsky (UNN)

Doctor of Philosophy - PhD

2001 — 2006

State University of Nizhni Novgorod named after N.I. Lobachevsky (UNN)

Specialist Degree

SKILLS

EclipseWxwidgetsMapreduceXmlArtificial Neural NetworksPandasParallel ProgrammingNeural NetworksComputer ArchitectureNaive BayesSubversionPerformance AnalysisStlMongodbC++Information RetrievalScikit-LearnVisual StudioDebuggingConvolutional Neural NetworksObject Oriented DesignUmlRnnPrologGitGradient BoostingHaskellDeep LearningProfiling ToolsTest Driven DevelopmentPinDatabasesNumpyCLtsmPythonComputer ScienceLatex.netMultithreading

ABOUT ANTON DERGUNOV

Machine Learning: Extensive experience and deep interest in applying machine learning…

This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.

Anton Dergunov — Senior Applied Scientist at Microsoft in London, GB | Unifers