• Cut ML operational costs by 60% while maintaining availability, by strategically managing GPU nodes and adopting container image prefetching, leveraging deep Kubernetes and AWS expertise.• Increased LLM inference throughput by 3 times, by optimizing models and offloading inference storage.• Built event-driven distributed compute system, handling >10x LLMs, processing millions of requests daily.• Reduced backend response times by 50% by rewriting hot code with async Python.• Scaled infrastructure from single-user POC to supporting hundreds of users, and in-house ML experiments. • Built in-house MLOps, enabling hundreds of daily deployments and reducing build-to-deploy times by >10x.

- Building real-time ML systems, by designing fast and scalable web APIs, and optimizing models for hardware utilization.- Leading development of low-code customizable virtual assistant, building system with integrated MLOps pipelines for training and updating state-of-the-art models.- Supervising project to optimize bidding, exploring use of supervised learning with advanced techniques like SentenceBERT and metric learning.- Built image processing algorithm for detecting obstacles on embedded cameras. Scaling application to use with distributed systems containing network of cameras.

- Built, trained and deployed ML models for real-world applications.- Led project to develop recommendation model, reducing client’s operational costs by 15%. - Designed and built real-time object detection model, attaining over 900 FPS with sub-second latencies. - Added ASR and NLP models into client’s legacy system, reducing operational resolution times by up to 90%. - Initiated and maintain the team’s MLOps infrastructure, accelerating deployment cycles by 200%. - Awarded company top performer 2021.

- Built backend for handling terabytes of geographic information and map data.- Initiated and migrated GUI application from GTK to Qt, accelerating product’s cross-platform rollout.