I'm Nikita — I train models, and then I build everything that has to exist around them so they actually run for somebody other than me.

BSc in Computer Science, Don State Technical University · Russian Federation. Python most days.

Speech, emotion, and voice. Most of my research time goes into audio and multimodal models — recognising emotion from speech in real time, optimizing speech recognition in real time and synthesising Russian speech that doesn't sound like a 2015 TTS demo. This is the part of ML I keep coming back to: the data is messy, the labels are subjective, and the evaluation is genuinely hard. Fine-tuning transformers on it is the easy half.

LLM pipelines that have to survive contact with production. OCR in front, an LLM in the middle, a document that a real business depends on coming out the other end. The interesting problems there aren't prompts — they're retries, idempotency, queue backpressure, versioning the model and the data together, and what happens at 3 a.m. when the extraction quietly starts returning garbage.

The unglamorous infrastructure underneath. Async APIs, brokers, containers, deployment. I like that a model is only as good as the boring plumbing that keeps it fed and observable, and I'd rather own that plumbing than hand it over and hope.

• Aniemore — an open library for real-time speech emotion recognition using a multimodal approach. Audio and text together, because tone alone lies and words alone lie differently.
• OpenJourney — a Discord bot wiring Stable Diffusion to a GPT-2 I trained specifically to write image prompts. Turns out the hard part of image generation is asking for the right thing.

• Document Automation System — an OCR + LLM pipeline processing documents end-to-end for a logistics company. Real throughput, real consequences for getting a number wrong.
• Chtec — a speech-synthesis prototype built on StyleTTS 2 for voice cloning in Russian. Cloning a voice in a language the model wasn't designed around is where most of the work went.

PyTorch · HuggingFace (transformers / datasets / accelerate / peft) · pandas · polars · sklearn · lightgbm · catboost · langchain · gigachain · langfuse · plotly · seaborn FastAPI · Pydantic · FastStream · RabbitMQ · Redis · PostgreSQL · MySQL · minio Docker · Compose · Swarm · Ansible/Semaphore · DVC · k8s · ArgoCD · Helm · vLLM · GPUStack

Tools are the least interesting thing on a profile, so that's all they get. I pick them up when a problem needs them and I'm happy to learn whatever the next problem needs — especially if it comes with large-scale infrastructure attached.

Pinned repositories are below. Thank you and enjoy ❤️