LLM · multimodal models · real-world AI
Nikolai
Zakharov
I build transformer systems from data and distributed training through evaluation and production inference—across LLMs, generative models, autonomous-driving planning, and speech.
- 8+ years
- ML research & engineering
- Full stack
- Data · training · inference
- Current direction
- Multimodal & embodied AI
Current focus
Models that hold up from training to real-world evaluation.
My work connects model quality with the systems required to train, evaluate, and run models reliably. I am applying that experience to multimodal and autonomous systems while moving deeper into embodied AI.
What I work on
LLM & multimodal training
Fine-tuning, distillation, synthetic data, distributed training, reproducible datasets, and experiments that connect research ideas to usable models.
Evaluation & inference
Multimodal and geometric evaluation, open/closed-loop metrics, TensorRT, streaming services, and production inference across GPU and edge systems.
Autonomous & embodied systems
Transformer planning, 3D generative modeling, closed-loop evaluation, and model systems that transfer toward real-world embodied AI.
Selected experience
Eight+ years across ML research and industry.
2025—26
AI Startup · VK · Kandinsky Lab (Sber AI)
Dental-crown generation, music-video LLM captioning, recommendation evaluation, Qwen prompt enhancement with LoRA, and multi-GPU evaluation.
2023—25
NIO
End-to-end transformer planning and C++ rule-based failure detection integrated into open/closed-loop evaluation.
2019—23
T-Bank · Huawei
Voice conversion for a 10M+ MAU assistant, multilingual NLU, shipped on-device voice cloning, and TensorRT/TFLite/ARM NEON optimization.
Notes & projects
Work in public.
Research note · 23 min
From Sound to Meaning
How audio language models can connect acoustic similarity with perceptual relevance in music recommendation.
Read articleOpen source project
Paper Clusterer
An Obsidian plugin that turns research notes into a navigable knowledge map with embeddings and AI-assisted labels.
Explore projectSay hello