Built from scratch or rebuilt from something that wasn't working. Newest first.
The closing iteration of my serverless lakehouse: an AWS twin of the same platform, built from the same data and the same models. The Yandex Cloud original keeps running beside it. S3, Spark on AWS Glue, Iceberg tables, the Glue Data Catalog, Athena, Secrets Manager, Budgets and Cost Anomaly Detection, all in Terraform. I learned AWS by building this. The story is what changed moving from DuckDB and pointer files to Glue and Iceberg.
One analytics platform for a bookstore chain, rebuilt four times as the workload taught me what it needed. First PostgreSQL with Airflow in Docker on Yandex Cloud and DataLens dashboards. Then Airflow with SQLMesh models, four Postgres instances and Superset: a proper transformation layer, on a large always-on VM for a small workload. Then a serverless, scale-to-zero lakehouse: Parquet on object storage, DuckDB for every transformation, YAML data contracts that validate each dataset and generate the catalog, atomic publish, all Terraform, at about 5× lower TCO than the VM it replaced. The fourth iteration is the AWS twin. Finished in 2026; it runs unattended.
Re-architected the company data platform on a hybrid S3 + ClickHouse medallion design, with analytical queries served directly on low-cost object storage. Turned a platform with one to two bad-data days a month into the trusted source of truth: SQLMesh audits and tests drove data-quality incidents to near zero, where they have held for a year. Standardized pipeline deployment, so onboarding a new source went from hours to minutes. Deployed OpenMetadata as the company-wide catalog and wrote the catalog standards that other teams now follow.
Open WebUI as the shared interface for open-source LLMs, vLLM for inference with observability, and a RAG system built for legal-document search. Internal services are getting MCP interfaces, with Hermes, an open-source agent, as the main agent analysts work with: one chat that reaches the whole stack.
The six-stage loop I work in, packaged as a Claude Code plugin so it is fixed in one place instead of re-derived in every repo: intent, spec, plan, build, review, watch. The spec stage is BDD. Every behavior becomes a Gherkin scenario in a .feature file, and that file is the contract the build is written against and reviewed against. A hook blocks edits to an existing scenario, so the target cannot move quietly to meet the code. Project-agnostic, with no language or stack assumption in it.
Designed the ClickHouse-based data-mart layer of a data warehouse and built Spark and Airflow ETL pipelines for analytical and ML workloads.
Built an automated analytics system — PostgreSQL with Airflow in Docker on Yandex Cloud — and DataLens dashboards that surfaced weak points in inventory management.
A Dask/NumPy and Hadoop processing pipeline with a PyTorch model for radar data, lifting F-score from 0.25 to 0.65 at the Institute of Radar Meteorology.