Gleb Tkachev
ML Scientist
Zalando
About me
I am a Machine Learning Scientist working on methods that solve practical problems. My current position is at Zalando, where I develop RecSys models and LLMs that help people find well-fitting clothes. Before that I was at Amazon Alexa, working on ranking for Question Answering. And prior to Amazon, in my PhD, I worked on ML for rendering and analysis of scientific data. I focused on self-supervised representation learning and developed interactive performance-optimized methods. Additionally, I have solid software development skills and familiarity with performance optimization and distributed systems.
-
Machine Learning
- LLMs & NLP
- Recommender Systems
- Self-Supervised Learning
-
Software Development
- Python (Pytorch, Pyspark, Numpy)
- C++ (Pybind, CUDA)
- Performance optimization
-
PhD Computer Science, 2022
University of Stuttgart
-
M.Sc. Computer Science, 2017
University of Stuttgart
-
B.Sc. Automatic Systems, 2014
Kaliningrad Technical
Experience
Working on recommendation models that help people find well-fitting clothes.
- Proposed a new DL recsys model that significantly outperforms the production model. Brought the project from conception to experimentation and finally to production.
- Developed an LLM-based recommendation approach that outperforms DL methods. Built a customer dialog experience with LLMs. Tuned open-source LLMs, built automated eval.
- Significantly improved the team’s codebase, data pipelines and the overall approach to operational excellence.
- Mentored junior colleagues on machine learning and code quality topics.
Working on Answer Ranking models for a Question Answering system.
- Delivered customer-relevant model improvements within a few months of onboarding to the team.
- Automated offline model experiments, significantly reducing the time wasted on experiment orchestration.
Researching Machine Learning approaches to visualization and analysis of scientific data (see publications).
- Introduced a self-supervised similarity metric for scientific data. The metric outperformed alternatives and helped the domain scientist discover a new feature. Published in a top graphics journal – TVCG.
- Developed a NN-based anomaly detection approach for spatiotemporal volumes. Found anomalies missed by several domain experts. Published in TVCG.
- Built a distributed volume renderer~(C++, CUDA, MPI). Proposed an ML model for performance prediction.
- Built and managed a GPU cluster to handle ML experiment workloads (Linux, Slurm, Singularity).
- Supervised six Master and Bachelor student theses, as well as seminars and student research projects.
A part-time web-development job during my Master studies (20h/week).
- Created web-based business applications and development tools. (PHP + CSS + Javascript)
- Linux administration. (Debian + Apache + Freeswitch)
A part-time web-development job during my Master studies (20h/week).
- Developed an information visualization prototype. (Javascript + D3)
- Mobile MMORPG server development. (C# + .NET)
- Created a compiler for an in-house networking solution. (C# + .NET + Irony)
- Mobile game development, rapid prototyping. (Unity3D)
- Web-based business application development. (PHP + Yii, PostgreSQL, Javascript + Knockout)
- Web-server administration. (Ubuntu server, Apache, Postfix, Samba)