Nikolai Zakharov

Nikolai Zakharov (罗一阳)

Senior Applied ML / GenAI Engineer working across LLM, multimodal, generative, and autonomous systems.

I build transformer systems from data and distributed training through evaluation and production inference. My work spans LLM distillation and synthetic data, 3D generative modeling, autonomous-driving planning, closed-loop evaluation, and efficient speech systems. I am now focusing that experience toward embodied AI, where model quality must hold up in real-world feedback loops.

Core expertise: Transformer training • Multimodal evaluation • Production inference • Autonomous systems

Quick Facts

  • Languages: Russian (native); English and Mandarin Chinese (fluent); French and Turkish (basic)
  • Current focus: LLM and multimodal training, evaluation, inference, and embodied-AI model systems
  • Experience: 8+ years spanning academic ML research and industry
  • Education: Tsinghua University, M.S. in Computer Science (Machine Learning)
  • Location: Shanghai, China; open to worldwide remote work and relocation

My Journey

The common thread in my work is taking learned models into systems where quality must survive real constraints:

  • Huawei: multilingual NLU; shipped on-device voice cloning and contributed to its on-device training framework with Huawei Russia; researched heterogeneous on-device RNN-T ASR; optimized on-device perception with ARM NEON.
  • T-Bank: transformer voice conversion in a 10M+ MAU assistant product, Go streaming inference, and TensorRT optimization.
  • NIO: end-to-end transformer planning plus C++ rule-based failure detection integrated into open/closed-loop evaluation.
  • VK and Kandinsky Lab (Sber AI): music-video LLM captioning, distillation, recommendation-quality dashboards, Qwen prompt enhancement with LoRA, synthetic data, and multi-GPU evaluation.
  • Generative-model startup: dental-crown generation models and supporting multi-GPU/Kubernetes infrastructure.

This path—from language and inference to planning, geometry, and closed-loop evaluation—is why I am moving toward embodied AI through model training, evaluation, and deployment systems.

What I Do Today

  • Train and fine-tune transformer, LLM, multimodal, and generative models
  • Design reproducible datasets, experiments, and evaluation systems
  • Build efficient inference paths across Python, C++, Go, TensorRT, GPU clusters, and edge devices
  • Connect model metrics with real-world and closed-loop behavior

Let’s Connect

I am interested in senior engineering and applied-research work around LLM and multimodal systems, model evaluation and inference, autonomous systems, and embodied AI.

LinkedIn Telegram GitHub

Collaboration & Speaking

I’m open to:

  • Technical writing and talks on model training, evaluation, inference, and multimodal systems
  • Open-source contributions in LLM, robotics, and ML systems
  • Applied research collaborations around autonomous and embodied AI
  • Advisory work for teams taking ML models from research into production