Zygimantas Straznickas

Independent ML research, focusing on - LLM performance engineering, both for training and inference - Investigating counterintuitive behaviors of RL-tuned LLMs, including RLHF and more standard, non-human-feedback-based reinforcement learning environments - Winner of the NeurIPS 2023 Trojan Detection Challenge, Trojan Detection Track

Helped kickstart the company. Focus on all research/technical aspects, mainly AI and PL.

- Worked on the full LLM stack — training data collection and preprocessing, distributed pretraining, fine-tuning and inference. - Implemented custom distributed training code to accommodate a nonstandard GPU network topology.

• Worked on Bigtable, Google’s internal non-relational database system. • Designed and implemented systems to monitor silent data corruption in Bigtable, find corruption patterns and minimize the frequency of corruption incidents. • Led several deep drill-down analyses into the system’s concurrency logic, removing performance bottlenecks and fixing long-standing correctness bugs.

• Maintained and improved a statistical pipeline for measuring incoming data quality. • Designed and implemented new tools for data verification, combining automatic scraper-based techniques and manual human-guided judgements. • Designed and implemented a machine learning pipeline to predict the quality of incoming data, detect drops in accuracy and quickly react to them.

• Worked on developing a distributed word embedding learning framework. • Implemented a scalable word embedding evaluation system in Spark.

Maintenance and improvement of existing web projects and systems.

A food/recipe website with social network elements. I mostly focused on back-end development using PHP (Laravel framework), MySQL and Redis.