DL Developer @ GigaChat.Audio
SberDevices
SberDevices develops smart devices and AI services within Sber, one of Russia's largest banking and technology groups.
Dec 2024 — PresentTrained AudioLM and built large-scale audio data pipelines, producing over a million labeled hours and increasing TTS throughput eightfold while reducing cost per audio hour.
• Trained an AudioLM model across music, speech, and environmental sound domains, covering ASR, sound/music captioning, gender/age/language classification, sound/music genre multilabel classification, emotion recognition, diarization, and speaker count.
• Outperformed the baseline cascade system (ASR encoder + LM) across multiple benchmarks, including MMAU (MMLU for audio), reaching 0.78 on three target domains while preserving the capabilities of the specialized ASR model GigaAM (WER < 0.05 on Russian, < 0.13 on English).
ML Developer @ Yandex.Maps
Yandex LLC
Yandex is a major Russian technology company with products in search, maps, transportation, e-commerce, and cloud services.
May 2024 — Jul 2024Improved Yandex.Maps photo-intent classification, raising long-tail recall by about 43% and overall recall to 35% with a two-tower model.
• Worked on large-scale multi-label photo classification pipeline: given a user photo and a fixed, periodically expanding vocabulary of place intents (“alcohol”, “japanese cuisine”, “interior”, “atmosphere”, etc.), predicted which intents are present, across millions of photos.
• Rebuilt the online evaluation pipeline that had been degrading due to stale permalinks: designed a stratified sampling scheme over the (photo, intent) pairs with text-similarity-based threshold calibration to keep positive rates usable for annotation.
Teaching Assistant
NRU Higher School of Economics
The Higher School of Economics is a major Russian university, known for programs in economics, social sciences, and computer science.
Sep 2021 — Jun 2022Held classes and graded work for Machine Learning-1 and Math Statistics courses at HSE.
• Faculty of Computer Sciences, "Machine Learning-1" course
• Faculty of Economic Sciences, "Math Statistics" course
Development of educational materials
Karpov.Courses
Karpov.Courses is an online education platform offering courses in data analysis, machine learning, and related technical fields.
Mar 2022 — Jun 2022Created practice notebooks for participants in a Classic ML course.
• Making practice notebooks for Classic ML course participants
ML Engineer (Yandex School of Data Analysis)
Yandex Brain Segmentation Project
A Yandex School of Data Analysis project applying machine learning to MRI segmentation for diagnostics of newborn brains.
Trained a 3D newborn brain MRI segmentation model that outperformed the BIBSNet baseline; results were published in a Yandex Habr write-up.
• Trained the main 3D MRI segmentation model for newborn brain diagnostics, predicting white/grey matter masks on scans of infants under 12 months: adapted the PyTorch Lightning pipeline to 3D convolutions, integrated the iSeg-2019 dataset, and ran the full experiment pipeline across architectures (U-Net, U-Net++, DeepLabV3), backbones (ResNet-50/101, ResNeXt-50/101, DenseNet-161), and losses (Dice, Dice+CE).
