Vladislav Kruglikov

• Created distributed inference deployment for open source LLMs that allowed to use LLMs with confidential information and collect data with the quality compared to proprietary models in the cloud by serving LLMs with over 1 trillion parameters on premise cluster

• Lead a small team to built natural language processing platform that cut trained model TTM from 1 day to 1 minute with the same model quality by training adapters with parameters efficient fine tuning and building custom inference server runtime on top of nvidia triton inference server

• Built cluster analysis platform with user interface that allowed analysts to get insights from data 1000% faster with 90% better SBS quality by doing cluster analysis on GPU and training domain specific embedders

• Built personal data detection system that allowed 1000+ people to use proprietary LLMs on the cloud to solve tasks by verifying that requests do not contain company users data before sent to the cloud

• Built a system to train and evaluate named entity recognition models that allowed developers to save 400% of development time by leveraging reusable components for both training and evaluation and achieved 100% test coverage across all components

• Built a system for aspect sentiment triplet extraction that achieved 90% recall and 80% precision by collecting high quality data through internal crowdsourcing platform and implementing active learning pipeline