ALEXANDER NAGAEV
]LinkedIn − aGitHub − ûTelegram − Website : @nagadit
PROFILE
ML R&D Tech Lead with 5+ years of expertise in NLP and Computer Vision, dedicated to advancing state-of-the-art
AI through novel algorithms and architectures. Transitioned from foundational NLP research (2 years) to pioneering work in
Computer Vision, designing SOTA models for solving complex theoretical and applied challenges.
EXPERIENCE
Computer Vision Tech Lead June 2020 − Present
SberDevices Moscow, Russia
• Research
Papers, SOTA models
– The world’s 1st open source framework for recognizing isolated Russian sign language SLOVO. Achieving SOTA results
on benchmarks with 65.3% accuracy on WLASL (American sign language) and 87.3% accuracy on SLOVO. Conducted
in-depth research, culminating in a published paper presented at the in the ICVS’23 and CVPR’23 workshop.
– Developed an end-to-end algorithm for generating a real-time talking-head assistant/chatbot for video conferencing,
integrating 5 key modules: a spotter for assistant activation, an ASR-API for streaming voice-command transcription,
RuGPT-3 for response generation, a proprietary text2landmark model for lip-sync animation, and a head animation
system.
• Development
Created unique algoritms and applications
– Algorithm for recognizing dynamic gestures for devices, leveraging heuristics and hand detection models. The solution
enabled precise gesture-based interaction, enhancing user experience for device control applications.
– Demo stand for a Russian and American Sign Language teacher, featuring models capable of recognizing over 1,000
gestures in real-time on CPU. The stand was showcased at international conferences ICTWEEK’24 and GITEX’24,
highlighting its practical application and technological innovation.
– Real-time facial enhancement system for CPU environments, integrating 3D morphing (cheek/eye/jaw transformation
via MediaPipe landmark triangulation and Grid Sample mesh deformation), adaptive skin smoothing with Guided Filter,
and dynamic eye color modication. Optimized performance to 30 FPS on mid-tier CPUs through lightweight SSD
face detection and OpenCV pipeline enhancements.
– End-to-end pipeline for generating a real-time talking-head assistant/chatbot for video conferencing, integrating 5 key
modules: a spotter, an ASR-API for streaming voice-command transcription, RuGPT-3 for response generation, a
proprietary text2landmark model for lip-sync animation, and a head animation system.
Data Analyst October 2019 − June 2020
Active Business Consult Moscow, Russia
• Designed an error correction system for dialogue interfaces using synthetically generated ASR errors, achieving 15% WER
reduction via a context-aware ELMo model optimized for 512-token sequence processing.
PUBLICATIONS AND PRESENTATIONS
Science :
• Google Scholar
• Orcid
• Scopus
Articles
• Habr
Technical Talks
• Saint HighLoad++ 2024 lYouTube
• GIGA RnD Day 2024 lYouTube
• HighLoad++ 2023 lYouTube
• X5 Data science meetup 2023 lYouTube
EDUCATION
Master of Data Science, Ural Federal University 2019 − 2022
Instructed undergraduate students in Machine Learning and Python programming for nal-year bachelor’s courses while deep-
ening my own data science expertise.
Bachelor of Computer Science, Ural Federal University 2015 − 2019
Led cross-functional teams in semester-long academic projects using Agile methodology: Conducted sprint planning and retro-
spectives, Managed workows via Trello/Kanban boards and Collaborated in iterative development cycles
LANGUAGE
Russian (Native), English (A2 - B1)
