Maxim Sokolov

Delivered a realtime LLM-based chatbot supporting 10+ languages and 20+ countries, enabling 10k calls/day and achieving 95% quality by LLM-as-a-Judge; built auto-scaling production up to 20+ nodes and led dataset labeling and benchmarking efforts.

Led ML team to optimize AI-generated avatar pipeline, fixing lip jitter, improving smoothness, achieving average FID ~12 and reducing avatar generation time from days to 2 hours while managing ML and backend engineers.

Built and shipped AI-generated reels and image captioning products as the first employee, produced multiple prototypes weekly, and owned research, development, and ML infrastructure for production.

Delivered multiple ML solutions including sentence-matching with 80% top-1 precision, published comparative LLM research, automated MLOps pipelines with SageMaker, and launched 5 production ML solutions.

Developed custom video-analysis solutions for safety use-cases, doubled inference throughput by migrating to Triton, improved accuracy 10% via model architecture changes, and introduced reproducible experiments with DVC and MLFlow.

Led R&D to production for traffic flow video analytics, launched 1500+ edge boxes, raised core metric quality from 70% to 90%, formed four product lines, and secured partner credits from NVIDIA and Intel.

Activities and societies: Startup-focused learning program with courses from ETH Zurich, EPFL, IMD, and Constructor University; thesis on LLM development (RAG).

Master's program in Computer Science and Software Engineering with coursework from partner institutions and a thesis on LLM development (RAG).

Activities and societies: Coursework in algorithms, software engineering, AI and robotics; thesis on development of software for a crawler-type mobile robot.

Bachelor of Computer Science covering algorithms, OOP, databases, OS, cloud computing, software architecture, AI, and a thesis on software for a crawler-type mobile robot.