Dmitry Kulbida
ML Engineer
4+ years exp
4+ years ML exp
C++
Bash
C
Python
SQL
Active 17 days ago
+7 (919) 5578690 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A practical ML engineer at a middle level who produces hands-on notebooks, custom algorithm implementations and standard Transformer fine-tuning pipelines. The strongest proven skill is end-to-end model prototyping for classification tasks, supported by the BERT fine-tuning notebook (train_epoch/eval_model, DataLoader/tokenizer) and multiple tabular-model notebooks with preprocessing and evaluation. There is little to no evidence of production-grade MLOps, experiment tracking, deployment, or novel research in public code.
Technical skills
C++
Bash
C
Python• Middle • 4y+
SQL
C++
PyTorch C++
TensorFlow C++
Python
Flask
FastAPI• 4y+
aiogram
SQLAlchemy
Asyncio
Celery
Pydantic
Databases
FAISS
SQLite
ElasticSearch
Chroma• 4y+
PostgreSQL• 4y+
RabbitMQ• 4y+
Redis• 4y+
Weaviate• 4y+
OpenSearch
Qdrant
MinIO
AI/ML
Llama
Mistral
Mistral SDK
NLP
Perplexity
PyTorch
Semantic Search
TensorFlow
NumPy
Pandas
Deep Learning
BERT
CatBoost
Classic ML
Tokenizers
Embeddings• 4y+
Fine-tuning• 4y+
LLM• 4y+
LoRA• 4y+
MLFlow• 4y+
Quantization• 4y+
RAG• 4y+
Transformers• 4y+
PEFT• 4y+
Prompt Engineering• 3y+
QLoRA• 3y+
DeepSeek
GigaChat
LangChain
Langfuse
LangGraph
Scikit-learn
YandexGPT
DevOps
Git
Vector
CI/CD• 4y+
Docker• 4y+
Docker Compose• 4y+
Rest API• 4y+
Grafana
Prometheus
Kubernetes
VictoriaMetrics
Analytics
Matplotlib
Seaborn
QA
Pytest• 4y+
Frontend
GraphQL
Timeline
Tech Lead
•
Lead
Severstal
•
Full-Time
Led development of an AI/ML platform for market intelligence and patent competitive analysis. Designed a FastAPI-based service architecture with search components, background workers, and a clear internal/external integration boundary. Built a multi-source research pipeline with parallel search and deduplication, and implemented a patent RAG retrieval flow using lexical and vector search with scoring. Set up containerized infrastructure and data stores and added automated tests and technical documentation.
FastAPI
Pydantic
SQLAlchemy
PostgreSQL
Redis
Celery
Docker
Docker Compose
Embeddings
Qdrant
OpenSearch
MinIO
Rest API
RAG
LLM
Novosibirsk State Technical University
Bachelor's Degree •
Software Engineering
ML Engineer
•
Middle
MTS
•
Full-Time
Developed a multi-agent AI system that analyzes production incidents and supports postmortem/RCA reporting. Implemented pipelines to build incident timelines from logs, metrics, alerts, and related events, and used LLM/agent frameworks for reasoning and orchestration. Built backend services and APIs (REST/GraphQL), developed RAG components with multiple vector databases, and established monitoring/observability and CI/CD practices for production reliability.
LangChain
LangGraph
Langfuse
LLM
Embeddings
RAG
Qdrantsince 2025
Weaviate
Chroma
PostgreSQL
SQLAlchemysince 2025
Rest API
GraphQL
Docker
Kubernetes
Grafana
VictoriaMetrics
OpenSearchsince 2025
CI/CD
Fine-tuning
Prompt Engineering
Pytest
Transformers
SQL
ML Engineer
•
Middle
NCGI
•
Full-Time
Helped build and deploy a production Telegram bot using a multi-agent architecture with RAG. Owned technical architecture and integrations, from MVP planning through scaling and monitoring, including async task execution. Implemented backend and infrastructure components (FastAPI services, Redis, Weaviate) and message-driven processing with RabbitMQ. Integrated multiple LLM providers, built bot flows with aiogram, and set up containerization, CI/CD, and monitoring/alerting for stable operation.
aiogram
FastAPI
RabbitMQ
Redis
Weaviate
LangChainsince 2024
LangGraphsince 2024
Langfusesince 2024
RAG
LLM
GigaChat
YandexGPT
DeepSeek
Docker
CI/CD
Prometheus
Grafanasince 2024
ML Engineer
•
Middle
Online Gymnasium
•
Full-Time
Built and deployed an automated first-line support system and tools for tutors to reduce operational costs and improve customer onboarding. Collected requirements by analyzing business processes and defined automation scenarios and KPI-driven workflows. Implemented RAG assistants for support using corporate knowledge indexing and retrieval, and applied prompt engineering for more consistent model outputs. Developed FastAPI backend services with integrations to external systems and queues, added pytest-based testing and CI/CD, and supported production monitoring and prompt/knowledge updates.
FastAPI
RAG
LLM
Prompt Engineeringsince 2023
Pytest
CI/CD
ML Engineer
•
Middle
NutroTech
•
Full-Time
Designed and implemented an end-to-end document processing pipeline with preprocessing (including OCR), entity extraction, vectorization, and indexing. Developed and fine-tuned neural models for classification/regression and information extraction, and improved efficiency using quantization and LoRA/DoRA/QLoRA approaches. Built retrieval and RAG layers for relevance improvements, integrated LLM APIs for post-processing, and orchestrated asynchronous processing with RabbitMQ and Airflow. Implemented API services with FastAPI, added CI/CD and Docker-based deployments, and used MLflow along with pytest for experiment tracking and quality assurance.
Python
FastAPIsince 2022
Dockersince 2022
Docker Composesince 2022
RabbitMQsince 2022
CI/CDsince 2022
Pytestsince 2022
PostgreSQLsince 2022
Redissince 2022
Weaviatesince 2022
Chromasince 2022
RAGsince 2022
Embeddingssince 2022
LLMsince 2022
Fine-tuningsince 2022
Quantization
PEFT
LoRA
QLoRA
Transformerssince 2022
MLFlow
Rest APIsince 2022
Middle AI/ML Engineer
Confidence: Medium ML Engineer
A practical ML engineer at a middle level who produces hands-on notebooks, custom algorithm implementations and standard Transformer fine-tuning pipelines. The strongest proven skill is end-to-end model prototyping for classification tasks, supported by the BERT fine-tuning notebook (train_epoch/eval_model, DataLoader/tokenizer) and multiple tabular-model notebooks with preprocessing and evaluation. There is little to no evidence of production-grade MLOps, experiment tracking, deployment, or novel research in public code.
Model Architecture & Training
3/10
How well models are designed and trained
Implements models from scratch and standard Transformer fine-tuning loops (optimizers/scheduler/loss) but lacks custom architectures, ablations or rigorous training best-practices expected of senior work.
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Clear, hands-on preprocessing and feature engineering for tabular data and a tokenization/DataLoader pipeline for text classification; standard but practical ETL and dataset handling.
Evidence
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: preprocessing and feature creation (Age_Tenure_ratio, Age60+, ActiveUsers)
DomestosUltra/Classification-of-advertising-texts.-Fine-tune-BERT/TEST_KULBIDA.ipynb: ProductDataset and create_data_loader tokenization logic
Experimentation & Evaluation
3/10
How results are measured and tested
Uses GridSearchCV, cross-validation, ROC AUC and explicit train/eval loops to measure performance; however there is no experiment tracking, run management or structured ablation study.
MLOps & Deployment
1/10
How models are shipped to production
Minimal MLOps or deployment artifacts - only local model/tokenizer save and load; no serving, CI/CD, monitoring, or model versioning present.
Computational Efficiency
2/10
How efficiently computing resources are used
Some basic efficiency awareness (GPU device checks, DataLoader workers, n_jobs) but no evidence of profiling, quantization, mixed-precision or throughput/latency optimization work.
Evidence
DomestosUltra/Classification-of-advertising-texts.-Fine-tune-BERT/TEST_KULBIDA.ipynb: device = torch.device(...) and DataLoader(num_workers=4)
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: RandomForestClassifier(n_jobs=16)
Research Depth & Innovation
1/10
Depth of research and new ideas
Practical re-implementation of classic algorithms and a standard BERT fine-tuning pipeline but no original research, novel layers, reproduced paper results or rigorous innovation evidence.
Expertise
LLM• Middle
Industries
Financial Services• Middle
Technologies
Classic ML
SQL
C++
PostgreSQL• 4y+
Redis• 4y+
Weaviate• 4y+
LangGraph
Rest API• 4y+
LangChain
Docker Compose• 4y+
Chroma• 4y+
Flask
FAISS
Qdrant
DeepSeek
LoRA• 4y+
SQLAlchemy
RabbitMQ• 4y+
MLFlow• 4y+
MinIO
FastAPI• 4y+
Prometheus
VictoriaMetrics
Fine-tuning• 4y+
Embeddings• 4y+
Quantization• 4y+
Prompt Engineering• 3y+
NLP
Langfuse
Mistral SDK
PEFT• 4y+
QLoRA• 3y+
Llama
Mistral
YandexGPT
GigaChat
Tokenizers
CI/CD• 4y+
TensorFlow
Git
SQLite
PyTorch
Docker• 4y+
Kubernetes
ElasticSearch
Grafana
LLM• 4y+
RAG• 4y+
Perplexity
OpenSearch
Asyncio
Celery
TensorFlow C++
PyTorch C++
Pydantic
aiogram
Vector
Semantic Search
Recommendations
- Lead NLP classification and fine-tuning work - implement and optimize BERT-based pipelines and data loaders for text classification tasks.
- Prototype tabular ML workflows and feature engineering for business problems (Kaggle-style to production-ready transition), focusing next on reproducibility and validation.
- Invest in MLOps and reproducibility - add experiment tracking (W&B or MLflow), CI for training/saving artifacts, and simple model serving to move prototypes into production.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: Medium ML Practitioner
A machine-learning practitioner at a solid junior-to-middle level (tier_score 3.0) who produces working modeling notebooks and prototypes. The strongest proven skill is end-to-end model prototyping and training, evidenced by classical algorithm implementations and hyperparameter search (employee-churn-prediction.ipynb) plus a transformer-based fine-tuning pipeline (TEST_KULBIDA.ipynb). The public work lacks production-grade MLOps, comprehensive statistical rigor, formal testing and deployment automation.
Statistical Rigor
2/10
Correct use of statistics
Minimal statistical rigor - basic correlation checks and visual inspection are present but no formal assumption checks, uncertainty quantification, or hypothesis testing.
Evidence
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/employee-churn-prediction.ipynb: correlations = Churn_data.corr(); sns.heatmap(...)
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: correl = churn_data.corr(); sns.heatmap(data = correl, annot = True)
Data Wrangling & Cleaning
4/10
Preparing and cleaning data
Solid basic data wrangling - label encoding, scaling, feature creation and simple cleaning are implemented, but dataset provenance, detailed missing-value strategy and leakage checks are limited.
Evidence
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: label_encoder.fit(...); churn_data['Gender'] = label_encoder.transform(...); churn_data['Age_Tenure_ratio'] = churn_data['Age'] / churn_data['Tenure']
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/employee-churn-prediction.ipynb: MinMaxScaler applied to multiple numeric columns and LabelEncoder use on categorical columns
Exploratory Analysis & Visualization
3/10
Exploring and visualizing data
Exploratory visuals (histograms, KDE plots, heatmaps, boxplots) are used, but written interpretation and structured data-storytelling after plots are minimal.
Evidence
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/employee-churn-prediction.ipynb: Churn_data.hist(...); sns.boxplot(...); sns.heatmap(correlations, annot=True)
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: sns.kdeplot(...); churn_data.hist(...)
Predictive Modeling
4/10
Building models that predict
Shows practical predictive modeling skills - custom-from-scratch algorithms, GridSearchCV usage, CatBoost and a transformer fine-tuning pipeline, but lacks thorough error analysis, calibration and production concerns.
Evidence
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/employee-churn-prediction.ipynb: class LogisticRegressionFromScratch / class RandomForest / class KNN implementations and GridSearchCV usage
DomestosUltra/Classification-of-advertising-texts.-Fine-tune-BERT/TEST_KULBIDA.ipynb: custom DataLoader, train_epoch/eval_model functions and BertForSequenceClassification fine-tuning loop
Business Insight & Impact
1/10
Turning analysis into business value
Very limited business framing or impact analysis - predictions are produced and printed but there is no cost-of-error reasoning, KPI mapping, or stakeholder-focused recommendations.
Reproducibility & Notebook Hygiene
2/10
Clean, repeatable analysis
Some reproducibility steps exist (saving model/tokenizer, occasional random_state) but there is no pinned environment, experiment tracking, data versioning or modularized pipelines.
Evidence
DomestosUltra/Classification-of-advertising-texts.-Fine-tune-BERT/TEST_KULBIDA.ipynb: model.save_pretrained('bert_product_classification'); tokenizer.save_pretrained('bert_product_classification')
DomestosUltra/Binary-Classification-with-a-Bank-Churn-Dataset/binary-classification-with-a-bank-churn-dataset-n1 (2).ipynb: train_test_split(..., random_state=1)
Industries
Financial Services• Middle
Food & Beverages• Middle
Technologies
Deep Learning
Python• Middle • 4y+
CatBoost
Scikit-learn
Seaborn
Matplotlib
Transformers• 4y+
Pandas
NumPy
BERT
Recommendations
- Develop and productionize repeatable training pipelines - convert notebook experiments into modular scripts with config, fixed seeds and CI tests.
- Add experiment tracking and data versioning (MLflow/DVC or equivalent) and pin environment dependencies to improve reproducibility.
- Focus on deeper model evaluation - calibration, uncertainty, error analysis, and explicit leakage/assumption checks for each experiment.
- Prototype end-to-end deployment for one model (API + monitoring) to demonstrate operational ML skills beyond notebooks.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: