David Medvedev
ML Engineer
C#
Bash
Python
SQL
Model Architecture & Training: 4/10
Data Pipeline & Feature Engineering: 4/10
Computational Efficiency: 4/10
Active 6 days ago
Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Career analytics
Roles
Overview
A practical data engineer focused on building ETL orchestration and internal tooling with working knowledge of Airflow and data pipelines. The strongest proven skill is designing and implementing Airflow DAGs and simple custom operators as shown in airflow/plugins/postgres.py and the templated DAGs in airflow/dags/templs. What is not evidenced is mature production hardening such as automated testing, secure query parameterization, thorough performance tuning or formal model evaluation practices.
Technical skills
C#
Bash• Junior • 3y+
Python• Middle • 3y+
SQL• Middle
C#
.NET
Python
SQLAlchemy
FastAPI
Databases
MinIO
Qdrant
PostgreSQL• 3y+
AI/ML
NumPy
Embeddings
LLM
Prompt Engineering
PyTorch
Qwen
RAG
Reranking
Scikit-learn
XGBoost
Airflow
Pandas
Torchvision
Computer Vision
CatBoost
DVC
MLFlow
SciPy
Streamlit
vLLM
Frontend
Frontend
UI
DevOps
CI/CD
Grafana
Prometheus
Docker
Kubernetes
Rest API
Analytics
Matplotlib
Power BI• 3y+
Timeline
Data Scientist / ML Engineer
•
Middle
OOO Partner
•
Full-Time
Confirmed by code
80%
Разрабатывал внутренние ML/AI-системы для автоматизации процессов девелоперской компании. Создал локальную EngineeringRAG для работы с нормативной документацией: OCR-извлечение, гибридный поиск, reranking, сервис на FastAPI и интерфейс в Streamlit с offline-eval и проверяемыми ответами. Также построил ML-сервис прогнозирования годового бюджета с ETL и оркестрацией в Airflow, версионированием данных через DVC+S3, tracking в MLflow, и деплоем через Docker/Kubernetes; внедрил A/B-тестирование и статистические проверки для маркетинговых гипотез.
FastAPI
Streamlit
vLLM
Docker
Kubernetes
DVC
MLFlow
PostgreSQL
Power BI
SciPy
CatBoost
Python
SQL
Rest API
Tyumen Industrial University
Master's Degree
Moscow Institute of Physics and Technology
Master's Degree
Python Developer / Data Engineer, BIM automation
•
Middle
OOO Partner
•
Full-Time
Работал в BIM-команде над внутренними сервисами и задачами data engineering, переходя от BIM-специалиста к Python- и ML-направлению. Реализовал ETL-процессы для извлечения атрибутов из 3D/BIM-моделей в PostgreSQL и сделал проектные данные доступными для аналитики и BI-отчетности. Построил автоматический контроль качества 3D-моделей с сбором и очисткой данных, расчетом метрик и подготовкой датасетов для Power BI, а также автоматизировал файловые процессы между BIM-инструментами и внутренними системами.
Pythonsince 2023
Bash
PostgreSQLsince 2023
Power BIsince 2023
Career analytics
Verified work history and code activity in one view - open-source impact, work rhythm, trust signals and a forecast of when this candidate is likely to move.
2
Years active
214
Commits analyzed
Work rhythm
UTC+3
12 AM6 AM12 PM6 PM12 AM
Code activity
Low
Peak
Last activity
Aug 2026 · 0 Days
Active days (90d)
18
Typical work hours
8 AM - 4 PM
Open source & impact
17
Merged pull requests
Collaboration Quality
Middle AI/ML Engineer
Confidence: Medium ML Engineer
Computer vision ML engineer at a middle level specializing in end-to-end PyTorch training pipelines and data engineering integrations. The strongest proven skill is building and running a conventional CV training pipeline with mixed precision, checkpointing and submission generation as implemented in the XRay training scripts. There is limited or no evidence of production serving, automated experiment tracking, formal test coverage or distributed/quantized inference in public code.
Model Architecture & Training
4/10
How well models are designed and trained
Conventional model training and architecture adjustments for computer vision using PyTorch; includes a full training loop, checkpointing, mixed precision and standard loss/optimizer choices but no custom layers or novel architectures.
Evidence
Grand-X-Ray/Attempt 2/main.py: train_model function with training/validation loop, scheduler.step and loss aggregation
Grand-X-Ray/Attempt 2/main.py: model construction (models.efficientnet_b1) and classifier head replacement
Grand-X-Ray/Attempt 2/main.py: save_model and load_model checkpointing functions
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Practical data pipeline and ETL work using Airflow DAGs, custom Postgres operators and a PyTorch Dataset class for data loading and preprocessing.
Evidence
RevitETL/airflow/dags/templs/dag_custmers.py: Airflow DAG tasks performing merges and aggregations
RevitETL/airflow/plugins/operators/postgres.py: PostgresToDataFrameOperator and DataFrameToPostgresOverrideOperator
Grand-X-Ray/Attempt 2/main.py: XRayData Dataset class with CSV parsing and image transform pipeline
Experimentation & Evaluation
3/10
How results are measured and tested
Basic experiment and evaluation artifacts are present such as logging training/validation losses, plotting progress and creating Kaggle submission CSVs, but there is no experiment tracking, ablation or automated evaluation suite.
Evidence
Grand-X-Ray/Attempt 2/main.py: plotting of train_losses and val_losses and creation of submission.csv in create_submit
Grand-X-Ray/Attempt 2/main.py: train_model returns train/val losses for later analysis
RevitETL/airflow/dags/templs/dag_ml.py: train_model task that returns model and fit objects
MLOps & Deployment
3/10
How models are shipped to production
Operational glue is evident - Airflow orchestration, custom operators to move DataFrames to/from Postgres, and a Streamlit admin UI - but there is limited evidence of production serving, CI/CD, monitoring or model versioning.
Evidence
RevitETL/admin/app.py: Streamlit-based admin UI for managing projects
RevitETL/airflow/plugins/operators/postgres.py: DataFrameToPostgresOverrideOperator for persisting task outputs
Computational Efficiency
4/10
How efficiently computing resources are used
Some efficiency-minded choices appear in training and data loading such as mixed precision (autocast/GradScaler), scheduler use, pin_memory and num_workers for DataLoader, but no advanced profiling, quantization or distributed training evidence.
Evidence
Grand-X-Ray/Attempt 2/main.py: use of autocast and GradScaler for mixed precision
Grand-X-Ray/Attempt 2/main.py: DataLoader settings using num_workers and pin_memory
Research Depth & Innovation
2/10
Depth of research and new ideas
No custom research contributions or novel algorithmic work; the code implements standard models and external libraries rather than new layers, algorithms or reproduced SOTA experiments.
Evidence
Grand-X-Ray/Attempt 2/main.py: uses pretrained efficientnet_b1 and standard BCEWithLogitsLoss without custom layers
RevitETL/airflow/dags/templs/dag_ml.py: uses external ThymeBoost library rather than in-house algorithmic innovation
Expertise
Computer Vision & Image Analysis• Middle
MLOps & Model Lifecycle• Middle
Technologies
SQL• Middle
DVC
Rest API
Qwen
CatBoost
Qdrant
vLLM
MLFlow
XGBoost
MinIO
Prometheus
Embeddings
Scikit-learn
Prompt Engineering
SciPy
Computer Vision
CI/CD
NumPy
PyTorch
Docker
Kubernetes
Grafana
LLM
RAG
Torchvision
Reranking
Ray• mentioned only
Recommendations
- Develop productionizing skills for models: CI/CD, model versioning, and a serving stack (REST/gRPC endpoints with model checks and monitoring).
- Extend experiment rig: add experiment tracking (W&B/MLflow), reproducible configs, and clear evaluation/ablation scripts to validate model changes.
- Harden ETL operators and SQL usage: replace f-string SQL with parameterized queries and add input validation and unit tests for operators.
- Add lightweight profiling and efficiency targets: measure GPU memory/throughput, consider quantization or batch optimizations and document before/after metrics.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: Medium Data Engineer
A practical data engineer focused on building ETL orchestration and internal tooling with working knowledge of Airflow and data pipelines. The strongest proven skill is designing and implementing Airflow DAGs and simple custom operators as shown in airflow/plugins/postgres.py and the templated DAGs in airflow/dags/templs. What is not evidenced is mature production hardening such as automated testing, secure query parameterization, thorough performance tuning or formal model evaluation practices.
Statistical Rigor
2/10
Correct use of statistics
Basic model training workflow is implemented with train/validation split and loss tracking but there is no cross validation, uncertainty quantification, hypothesis testing or proper evaluation pipeline.
Evidence
Grand-X-Ray/Attempt 2/main.py: train_model
Grand-X-Ray/First attempt/main.py: train_model
Data Wrangling & Cleaning
4/10
Preparing and cleaning data
Multiple ETL DAGs, custom Airflow operators and a small config manager show practical experience with data ingestion, cleaning and Postgres integration, though input sanitization and large-table handling are weak points.
Evidence
RevitETL/airflow/plugins/postgres.py: PostgresToDataFrameOperator and DataFrameToPostgresOverrideOperator
RevitETL/airflow/dags/templs/dag_custmers.py: merge_orders_customers / fix_dates
RevitETL/admin/config.py: ConfigManager with _merge_defaults and _resolve_default_files
Exploratory Analysis & Visualization
2/10
Exploring and visualizing data
Visualizations are minimal and focused on training loss plots and simple per-class mean statistics without deeper EDA, interpretive writeups or structured insights.
Evidence
Grand-X-Ray/Attempt 2/main.py: matplotlib training progress plotting and printed class statistics
Grand-X-Ray/First attempt/main.py: training progress plotting
Predictive Modeling
3/10
Building models that predict
Solid end-to-end modeling code exists including a custom Dataset, training loop with AMP/GradScaler, scheduler and checkpointing, but there is no rigorous validation strategy, error analysis or baseline/comparison discipline.
Evidence
Grand-X-Ray/Attempt 2/main.py: XRayData, train_model, predict_test, save_model
Grand-X-Ray/First attempt/main.py: XRayData and training loop
Business Insight & Impact
1/10
Turning analysis into business value
Little evidence of business-metric framing, cost-of-error analysis or actionable recommendations tied to business KPIs.
Evidence
RevitETL/admin/app.py: project management UI (project CRUD) - shows product orientation but not business-metric linkage
Reproducibility & Notebook Hygiene
2/10
Clean, repeatable analysis
Some reproducibility primitives are present such as model checkpoint save/load and device detection, but there are no pinned environment files, seed handling, data versioning or CI/pipeline automation for repeatable runs.
Evidence
Grand-X-Ray/Attempt 2/main.py: save_model and load_model checkpointing
requirements.txt: dependencies listed (env not pinned)
Expertise
Analytics• Middle
Technologies
Python• Middle • 3y+
PostgreSQL• 3y+
Airflow
SQLAlchemy
Power BI• 3y+
FastAPI
Matplotlib
Pandas
Streamlit
Ray• mentioned only
Recommendations
- Develop and harden Airflow-based ETL pipelines and custom operators for Postgres-centric ingestion and transformation tasks.
- Build internal admin tools and lightweight UIs using Streamlit for project/config management and simple operational workflows.
- Implement end-to-end prototype CV training workflows that include evaluation pipelines, cross validation and error analysis.
- Improve production readiness by adding input validation, secure query parameterization, and CI/tests for operators and DAGs.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Frontend Developer
Confidence: Medium App Engineer
Frontend app engineer (middle) specializing in interactive, form-driven web applications for events and admin dashboards. The strongest proven skill is building client-side RSVP and preferences flows with robust payload shaping and restore logic as shown in static/js/main.js (buildPayload, collectDrinks/collectMenuChoices and sendRsvp). The public code lacks automated tests, structured request cancellation or advanced performance measurement and has limited a11y CI or focus-management for custom widgets.
UI Component Architecture
3/10
How interface parts are built
Vanilla-JS composition with many small DOM-focused functions (reasonable separation) but no component system or design-system boundaries; mostly one large imperative file.
Responsive & Cross-browser
4/10
Works on all screens and browsers
Mobile-first CSS and responsive breakpoints are implemented with sensible grid/flex fallback; basic meta viewport and simple accessibility hints present but no RTL/i18n strategy.
Performance Optimization
3/10
Speed of the interface
Some pragmatic performance-minded choices (IntersectionObserver for reveal animations) but no measured budgets, code-splitting, virtualization or bundle-analysis artifacts.
Accessibility & Semantics
4/10
Usable for everyone
Reasonable semantic markup and some a11y interactions (labels, Escape to close modal) but missing focus trapping, ARIA for custom tab panels and CI a11y checks.
State Management & Data Flow
3/10
Managing data in the app
Client-side state is handled explicitly with a small state object and careful payload building and restore logic, but lacks request cancellation, optimistic updates, cache invalidation or structured state machines.
UX & Visual Polish
4/10
Look and feel quality
Good UX touches such as smooth scrolling, toasts, auto-submit when possible and restore of saved selections; lacks skeletons, loading indicators for network actions, and undo patterns.
Expertise
PWA & Web APIs• Junior
Industries
Events & Ticketing• Middle
Technologies
Frontend
UI
Recommendations
- Implement a small state module or lightweight component boundary (even in vanilla JS) to split main.js into composable units and make edge-state testing easier.
- Add request cancellation/abort handling and UX loading indicators for network operations (use AbortController and explicit loading states around sendRsvp).
- Introduce automated tests for form flows and end-to-end scenarios (unit tests for payload building and an integration test covering submit/restore).
- Harden admin access patterns - avoid embedding passwords in rendered templates and switch to server-side session or token-based auth.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: