Evgeny Korovin
AI Engineer
Python
MLOps & Deployment: 4/10
Active 22 days ago
+7 (977) 3218479 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
LLM engineer at a senior level specializing in agentic workflow orchestration and API-driven ML services. The strongest proven skill is designing modular agent orchestration with retry/backoff and run-level logging as implemented in src/orchestrator.py and src/log_manager.py. There is limited evidence of rigorous experiment tracking, test coverage, GPU/efficiency engineering or custom research-model design.
Technical skills
Python• Senior
Python
Pydantic
Uvicorn
HTTPX
FastAPI
AI/ML
Claude
Cursor
LangGraph
LLM
OpenCV
RAG
XGBoost
Scikit-learn
Pandas
NumPy
Streamlit
AI Agents
OpenAI SDK
Prophet
Model Context Protocol
DevOps
Docker
Rest API
Analytics
Plotly
Timeline
AI Automation Software Engineer
•
Middle
Mikrolain Center
•
Full-Time
Develops production-grade AI automation solutions focused on connecting LLMs with workflow systems. Built an MCP gateway service using Python and FastAPI to orchestrate n8n workflows and expose tool endpoints for viewing, starting, activating/deactivating, and tracking runs. Works on business integrations involving REST APIs and Bitrix24/1C, supporting reliable AI-driven automation.
Python
FastAPI
LangChain
Model Context Protocol
Rest API
Bitrix24
Senior AI/ML Engineer
Confidence: High LLM Engineer
LLM engineer at a senior level specializing in agentic workflow orchestration and API-driven ML services. The strongest proven skill is designing modular agent orchestration with retry/backoff and run-level logging as implemented in src/orchestrator.py and src/log_manager.py. There is limited evidence of rigorous experiment tracking, test coverage, GPU/efficiency engineering or custom research-model design.
Model Architecture & Training
3/10
How well models are designed and trained
Practical applied model training with standard ML stack (XGBoost, scaler) but no custom architectures, distributed training or advanced tuning.
Data Pipeline & Feature Engineering
2/10
How data is prepared for models
Basic data preparation and synthetic data generation; no advanced feature engineering or leakage controls.
Experimentation & Evaluation
2/10
How results are measured and tested
Minimal evaluation (classification report, ROC-AUC printed); no experiment tracking, reproducible runs, or ablation setups.
MLOps & Deployment
4/10
How models are shipped to production
Evidence of serving and orchestration: FastAPI model serving and a modular multi-agent orchestrator with retry and run-logging, but limited production hardening or CI/CD artifacts.
Evidence
Computational Efficiency
1/10
How efficiently computing resources are used
No focused work on GPU optimization, batching, quantization or profiling is evidenced.
Research Depth & Innovation
1/10
Depth of research and new ideas
No research-level implementations, custom layers, or paper-reproduction artifacts; mostly application-level composition of existing models and APIs.
Expertise
AI Agents & Agentic Workflows• Senior
LLM• Middle
MLOps & Model Lifecycle• Middle
Industries
Artificial Intelligence• Middle
Technologies
Cursor
LangGraph
Rest API
Claude
OpenCV
Model Context Protocol
AI Agents
Docker
LLM
RAG
Recommendations
- Develop agentic orchestration for production-grade LLM workflows, including retries, cost accounting and run logs (build on src/orchestrator.py and src/log_manager.py).
- Implement experiment tracking and reproducible training pipelines (W&B/MLflow) and add test coverage for core components like the orchestrator and agents.
- Harden serving layers for production: add input validation, auth, structured error handling, and CI/CD deployment for FastAPI and MCP client integrations.
- Invest in performance and efficiency work for models and inference (batching, quantization, profiling) when moving beyond demo/demo-mode clients.
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
ML practitioner at a middle level specializing in end-to-end predictive systems that train models and expose them via FastAPI with Streamlit dashboards as the main UI. The strongest proven skill is building and wiring training-to-serving pipelines, evidenced by churn-prediction-ml/src/train.py training an XGBoost model and churn-prediction-ml/app.py serving predictions. Not evidenced are rigorous model evaluation workflows like CV and hyperparameter search, production-grade CI/CD, and large-scale distributed data engineering.
Statistical Rigor
3/10
Correct use of statistics
Basic evaluation is present (stratified split, ROC-AUC) but there is no cross-validation, uncertainty quantification, or formal assumption checks.
Evidence
churn-prediction-ml/src/train.py: uses train_test_split(..., stratify=y) and roc_auc_score
demand-forecasting-appliances/src/train.py: orchestrates model training and prints metrics
Data Wrangling & Cleaning
4/10
Preparing and cleaning data
Clear data generation and basic preprocessing (StandardScaler, saved scalers) are implemented, but there is limited explicit handling of missing data, provenance tracking, or leakage prevention beyond stratification.
Evidence
churn-prediction-ml/src/generate_data.py: synthetic data generation to data/raw/customer_data.csv
churn-prediction-ml/src/train.py: StandardScaler fit/transform and joblib.dump of scaler
Exploratory Analysis & Visualization
3/10
Exploring and visualizing data
Streamlit dashboards and simple visualizations exist and are interactive, but exploratory analysis lacks written interpretations, hypothesis-driven findings, or deeper EDA artifacts.
Evidence
churn-prediction-ml/dashboard.py: Streamlit UI with sliders and result metrics
demand-forecasting-appliances/dashboard.py: Streamlit + Plotly charting of predicted demand
Predictive Modeling
4/10
Building models that predict
End-to-end model training and serving are implemented (XGBoost, model persistence, FastAPI endpoints); however, there is little evidence of systematic hyperparameter tuning, robust cross-validation, calibration, or in-depth error analysis.
Evidence
churn-prediction-ml/src/train.py: XGBClassifier training, predict_proba and ROC-AUC reporting
demand-forecasting-appliances/src/train.py: sequential training calls for XGBoost, LSTM, Prophet agents
Business Insight & Impact
2/10
Turning analysis into business value
Some business-facing outputs exist (risk label, dashboards) but there is no documented connection to business KPIs, cost-of-error analysis, or decision thresholds tuned to business impact.
Evidence
churn-prediction-ml/app.py: returns 'risk' field based on probability threshold
demand-forecasting-appliances/dashboard.py: displays predicted demand with no business cost discussion
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Project structure, requirements and model persistence enable basic reproducibility and local deployment, but environments are not pinned, there is no data versioning, and there is no CI/pipeline automation shown.
Evidence
churn-prediction-ml/requirements.txt: dependency list
churn-prediction-ml/src/train.py: joblib.dump for model and scaler with random_state set
Expertise
Analytics• Middle
Industries
Commerce• Middle
Technologies
Python• Senior
XGBoost
FastAPI
Scikit-learn
OpenAI SDK
Prophet
Plotly
Pandas
NumPy
Streamlit
Pydantic
HTTPX
Uvicorn
Python• mentioned only
Recommendations
- Develop inference APIs and small-to-medium productionization tasks such as model serving, monitoring, and lightweight CI for ML endpoints.
- Implement systematic model evaluation workflows: add cross-validation, hyperparameter tuning, calibration and automated model comparison for reproducible results.
- Expand data engineering practices: add data versioning (DVC), pinned dependency files, and input validation to improve reproducibility and safety.
- Build richer business-aligned evaluations: define KPIs, cost-of-error analyses (FP/FN tradeoffs) and threshold tuning for production decisions.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Backend Developer
Confidence: Medium API Engineer
A backend API-focused Python developer (middle level) who builds API-driven AI and ML prototypes with clear separation of components. The strongest proven skill is integrating AI/LLM-style services and orchestration as shown by the ticket processor and orchestrator code (ai-ticket-processor/src/* and vibe-orchestrator/src/orchestrator.py). There is limited evidence of production-grade data persistence, schema evolution, structured observability or advanced security hardening in public code.
API Design
4/10
How well APIs are designed
Clear API-first work with FastAPI and pydantic models, use of response models and static file mounting, but no formal versioning, idempotency strategy, pagination or documented error contract beyond basic HTTPException handling.
Evidence
EvgenyLubitel/churn-prediction-ml/app.py: FastAPI routes / and /predict with Pydantic CustomerData model
EvgenyLubitel/ai-ticket-processor/src/api.py: FastAPI app with ProcessRequest/ProcessResponse Pydantic models and mounted StaticFiles
EvgenyLubitel/vibe-orchestrator/src/main.py: FastAPI app entrypoints (root, create_campaign, health, get_run_log)
Data Layer & Database
1/10
Working with databases
Minimal data layer; model persistence via joblib and simple JSON file writes. No database usage, no migrations, no transaction boundaries or tuned queries are present.
Evidence
EvgenyLubitel/churn-prediction-ml/src/train.py: joblib.dump('models/churn_model.pkl') and scaler persistence
EvgenyLubitel/churn-prediction-ml/src/generate_data.py: synthetic CSV generation to data/raw/customer_data.csv
EvgenyLubitel/ai-ticket-processor/src/agent.py: save_history writes data/tickets.json
Scalability & Performance
3/10
Handling load and speed
Some attention to network resiliency and client configuration (httpx.Client with timeouts, follow_redirects), and simple retry semantics in orchestrator, but no caching strategy, no connection pool tuning beyond httpx.Client defaults, and no measured performance optimization.
Evidence
EvgenyLubitel/vibe-orchestrator/src/mcp_client.py: httpx.Client creation with timeout and follow_redirects and _request wrapper
EvgenyLubitel/vibe-orchestrator/src/orchestrator.py:_execute_with_retry implements retry loop with sleep and attempt counting
EvgenyLubitel/vibe-orchestrator/src/vibe_client.py: httpx.Client usage for API calls
System Architecture
4/10
Overall system structure
Reasonable modular decomposition - orchestrator, agents, client wrappers and LogManager indicate deliberate separation of concerns and orchestration flow design, but services are in-process with no inter-service contracts or distributed system patterns.
Evidence
EvgenyLubitel/vibe-orchestrator/src/orchestrator.py: orchestrates Agent classes, MCP client and LogManager with run/campaign lifecycle
EvgenyLubitel/vibe-orchestrator/src/agents/base.py and agents/*: Agent abstraction and multiple agent implementations
EvgenyLubitel/vibe-orchestrator/src/log_manager.py: centralized run/step logging and lifecycle methods
Security & Auth
2/10
Protecting data and access
Basic security hygiene: configuration via environment variables and limited input validation through pydantic models, but no authentication, authorization, secrets management, token lifecycle, or explicit protections against common web threats.
Evidence
EvgenyLubitel/ai-ticket-processor/src/classifier.py: reads AITUNEL_API_KEY from environment and falls back to mock mode
EvgenyLubitel/churn-prediction-ml/app.py: Pydantic CustomerData used for request validation
EvgenyLubitel/vibe-orchestrator/src/mcp_client.py: reads VIBE_API_KEY and prints masked key, but no secret storage integration
Reliability & Observability
4/10
Stability and monitoring
Deliberate observability and reliability primitives exist: in-memory LogManager, run and step lifecycle, retries with status updates and timeouts on HTTP clients; however there is only print-based logging, no structured logs, metrics, tracing or alerting integrated.
Evidence
EvgenyLubitel/vibe-orchestrator/src/log_manager.py: RunLog and StepLog lifecycle management and methods to update steps and runs
EvgenyLubitel/vibe-orchestrator/src/orchestrator.py:_execute_with_retry updates step statuses and handles retries and failures
EvgenyLubitel/vibe-orchestrator/src/mcp_client.py and src/vibe_client.py: httpx clients configured with timeout and exception handling returning demo fallbacks
Expertise
Backend AI & LLM• Middle
Python• Middle
Industries
Artificial Intelligence• Middle
Data & Analytics• Middle
Sales & Marketing• Middle
Recommendations
- Develop API-driven ML and AI integrations and prototypes such as classification and summarization microservices with FastAPI.
- Implement agent orchestration and external service integrations, including retry/backoff and run-step logging for marketing or automation workflows.
- Build backend connectors to external AI vendors and robust client wrappers with improved error handling, metrics and secret management.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: