Dzmitry Shulhin

Specialize in LLM architecture and inference optimization using KV-caching, Mixture of Experts (MoE), and LoRA fine-tuning to maximize model efficiency. Build production-ready MLOps stacks with vLLM, Kubernetes, and MLflow for scalable, observable deployments.

Led cross-functional initiatives for high-load payment acceptance systems, orchestrating a platform TLS protocol upgrade while ensuring PCI compliance across microservices and maintaining high availability. Directed migration from embedded Hazelcast to a shared distributed cluster to improve resilience and observability while reducing infrastructure overhead and technical debt.

Re-engineered third-party agent availability tracking by replacing high-overhead MongoDB polling with Kafka broadcasting and local caching, improving real-time synchronization and reducing redundant database usage. Architected GPT-based agents and deployed ML-driven fraud prevention services, reducing false positives by 11%.

Worked on high-throughput data ingestion and downstream analytics reliability by implementing automated data cleaning and validation within Spark-based pipelines. Supported performance-focused ETL development to improve runtime efficiency for large-scale daily data processing.

Developed a Dockerized fine-tuning pipeline for Alexa custom voice generation using BERT and vocoders to improve speech synthesis quality. Improved AWS Polly in-memory cache hit rate by 18% and built Kinesis-to-S3 Lambda architecture pipelines for petabyte-scale data processing.

Engineered Spark-based ETL pipelines on Databricks, optimizing orchestration and performance tuning to reduce processing runtimes by 30%. Managed high-throughput terabyte-scale daily ingestion with automated data cleaning and validation for reliable downstream analytics.

Activities and societies: Building production-ready MLOps ecosystems using vLLM, Kubernetes, and MLflow for end-to-end AI lifecycle management with enterprise-grade observability and scale.

AI Performance Engineering Fellow specializing in LLM architecture and inference optimization, including KV-caching, Mixture of Experts (MoE), and LoRA fine-tuning.

Earned a Bachelor of Applied Science (BASc) in Computer Science at Belarusian State University.