Alexey Kuntsevich
Machine learning engineer, PhD Candidate in Physics
Summary
Data professional with 16 years of hands-on experience specializing in data science, machine learning, data engineering, and enterprise information system development. I work on ML-heavy projects since 2016 and for the last 3 years I’m focusing on language models. I’ve applied my in-depth technical knowledge to solve complex problems in insurance and digital marketing domains, using proactive and structured approaches. I have a strong track record in supporting teams through complete product life cycles, from requirements collection to model deployment. As a technologist, my primary focus is on employing cutting-edge data modeling and machine learning techniques to address intricate business challenges. I’m currently seeking a technical role to contribute to the development of the language model technology and engineering.
Last 10 years of job history
Inworld AI,
2024 - present
Staff ML engineer
- Development of state-of-the-art, latency-optimized information retrieval pipelines
- Large scale unstructured data analysis, curation of training and evaluation datasets
- Improving the quality of long-term human-computer interaction by implementing various memory mechanisms and representations
- Enabling internationalization support for RAG system
Allianz SE,
2021 - 2024
Language model engineer
- Fine-tuning models for specific domains: I’ve refined and fine-tuned language models to adapt to a particular area of knowledge and vocabulary, noticeably improving their performance and relevance.
- Training automation: I’ve designed and build an automated training&evaluation pipeline for Flan-T5/Flan-UL2 and LayoutLM models, to enable faster training&feedback cycles for the end-users.
- RAG enablement: with my help a team of data scientists managed to deliver RAG-based knowledge discovery and summarization application for the commercial insurance domain. I’ve provided efficient and scalable infrastructure and automation of the model serving, fine-tuning and data pipelining.
- Model integration and optimization: I’ve optimized language models for internal k8s-based serving, streamlining the integration process with various products and services. 4 applications went live to be actively used by stakeholders, more are in development.
- I’ve improved prompting for document Q&A specific applications to improve the quality of model responses.
- Code Optimization: I’ve taken the initiative to refine the code for T5 model family, particularly by enhancing encoder output caching and supporting larger context lengths. This has resulted in significant improvements in processing efficiency and model performance.
- LLM enablement: I develop documentation and materials to let the business stakeholder and technical teams decide for their path to integrate, serve and develop LLMs for their business cases. I run an internal LLM leaderboard, similar to one from huggingface for internal LLMs and benchmarks.
- Scalable model serving: I’ve introduced, maintained and monitored solutions for LLM serving, and gradually optimized them to increase throughput and flexibility of model serving.
2020 - 2021
Senior Data Scientist
- Gradually introduced a number of technologies and development approaches to enable team’s process and product scalability. That resulted in smaller development iterations and more reliable delivery of products.
- I drove smooth migration of data products from an on-premise infrastructure into the cloud while keeping the services uninterrupted.
- Developed an MLOps application framework which increased the team velocity by streamlining backlog.
Flixbus, Munich
2017 - 2020
Senior data scientist/Product owner, leading 6-people team in marketing automation domain.
- Scalable modeling and simulation: I introduced several data processing and simulation approaches to the team and led implementation of some of them (ad portfolio multi-actor simulations, bandit algorithms, massive cost-efficient data preprocessing for further modeling).
- AI-Driven Solutions: Together with the team, I designed and implemented a large-scale Ad portfolio forecasting and optimization framework, and integrated it into daily operations of digital marketing teams. This contributed to a 15% improvement of marketing ROI.
- Framework Implementation: I’ve introduced and continuously improved a process of reproducible experimentation of the ad portfolio optimization system. It enabled testing of the model improvements in isolation and on a small scale in live system, leading to shortening the model development cycle from several months to 2-3 weeks.
2014 - 2016
CHECK24, Munich
Data engineer, worked on the data warehouse processing insurance contracts data.
- Enabled a team of data scientists through ensuring availability, accessibility and latency of data in internal datasets. Supported a team of data scientists in building, deploying and integration of various models.
- Worked on a migration from Microsoft BI stack to OSS toolset (PostgreSQL, PyData) in order to provide better experience for data science team.
- Drove an end-to-end development of customer behaviour analysis model (including data acquisition, training process optimization and building an interactive visualization dashboard) that helped the business unit in understanding of customer behaviour, and journey, which paved the path toward experimentation and optimization of it, increasing conversion rates and improving customer experience.
Skills
Development stacks Python (poetry, fastapi, pytest), PyData stack (pandas, scikit-learn, seaborn, etc), Rust development stack (cargo, rocket).
Language modeling Information retrieval methods (RAG, vector search), Huggingface stack, T5x toolset, PyTorch, llama.cpp
Databases&storage PostgreSQL, MongoDB, Redis, RabbitMQ, Distributed storage (S3, Minio, Azure blob store)
OPS and Automation Docker, Ansible, Kubernetes, AWS, Debian family of Linux
Methodologies Agile from a perspective of a developer and product owner, scientific method, data-driven decision making, A/B experimentation, writing CV without usage of LLMs
Languages Russian (native), English (working proficiency), German (intermediate)
Just a fan of Vim, tmux, zsh, systemd, tabs, rust programming language, game theory
Education
2024
Professional courses
Systematically improving RAG applications
2012 - 2017
Nizhny Novgorod State university
PhD student on the topic of vacuum electrodynamics in particle accelerators.
Worked on particle accelerator electromagnetic field configuration improvement in order to achieve better electronic beam precision and focus by reducing harmful effects of beam oscillation.
2003 - 2009
Nizhny Novgorod State university
Diploma in physics on the topic of vacuum electrodynamics. MSc equivalent in Electrical Engineering (U.S. evaluation).
Certificates
AWS Solution architect - Professional
CSPO Certified Scrum product owner, 2019-2022.
MCP Multiple Microsoft certifications on SQL Server (development, administration, BI suite) and .NET development
PostgreSQL Certified PostgreSQL associate