Data Engineer and Data Analyst, looking forward into a Data Architecture, MLOps fields.

3+ years of BI analytical experience with 2+ years of hands-on engineering and programming experience within production environments, science, EU energy and e-commerce.

The European power grid, finally queryable. A 7-layer data platform turning 218,000+ FTP files into a queryable Apache Iceberg lakehouse with Trino SQL.

A solo-built, open-source data platform for ENTSO-E — the European electricity grid's transparency system. 218,000+ files across 89 publications, 94 unique schemas. Built from scratch: Python crawlers → PostgreSQL metadata catalog (74K records) → ephemeral S3 landing zone → PySpark MERGE INTO Apache Iceberg → dbt marts → Trino SQL.

• YAML-Driven SSOT Architecture: All configuration declarative: paths.yml for filesystem layout, publications.yml for 89 data sources. AST-based quality gates prevent hardcoding. Environment-aware (PROD/IOP).
• Delta Sync & Idempotent Loads: What was a 3-hour full FTP crawl is now a ~5-minute delta check. PySpark MERGE INTO Iceberg with stable keys — blow the pipeline away, rerun, same result.
• ADR-Driven Engineering: 10 Architecture Decision Records documenting every trade-off: YAML→PostgreSQL migration (ADR-008: 12.4 GB RAM → O(1) memory), ephemeral landing zone (ADR-004), dynamic schema inference (ADR-010).

Make Bayesian CNN research code face the EU AI Act. Industrial MLOps pipeline: train on GCP HPC, sign with Sigstore, deploy to real-time streaming inference — without editing a single line of the original research code.

A researcher wrote a Bayesian CNN for aircraft engine failure prediction. It works on his GPU. This repo makes it production-ready: CPU-only via adaptive runtime shim, SafeTensors instead of pickle, Sigstore keyless signing, and a 4-node real-time streaming pipeline (Redpanda → Bayesian inference → DuckDB → Streamlit dashboard).

• 14-point Adaptive Shim: Surgical monkeypatching layer intercepts CUDA calls, redirects I/O, replaces weak Bayesian defaults, parallelizes preprocessing — zero modifications to vendor source.
• Self-Destructing HPC Training: Terraform provisions a c2d-standard-32 GCE instance, Docker pulls the signed image, preprocessing runs in parallel across 32 cores, the Bayesian model trains, artifacts are sterilized → signed → provenanced → uploaded, and the instance deletes itself. Every run leaves a provenance.json birth certificate.
• Cloud-to-Edge Streaming Pipeline: NASA HDF5 telemetry → Parquet staging → Redpanda fleet simulator → Bayesian inference consumer (30-cycle sliding window, dual-domain Z-score recovery) → Streamlit Sentinel with LOWESS smoothing and real-time uncertainty (σ) visualization.
• EU AI Act Compliance: SafeTensors (zero RCE), Sigstore/Cosign (keyless attestation), KMS CMEK encryption (90-day key rotation), non-root containers (UID 1000), immutable provenance manifests, fail-closed forensic recovery on every exit path.

Expand 1 row in csv into 3 hours of an aircraft engine telemetry. Now do it for 100 aircrafts at the same time

This project simulates a fleet of aircraft engines generating high-frequency telemetry in real-time, ingests the massive event stream using a modern distributed stack, and delivers analytical health insights through a cloud data warehouse.

• High-Throughput Edge Simulator: Features a custom Golang simulator acting as an edge device, streaming millions of sensor telemetry records directly to Redpanda (Kafka).
• Structured Streaming & Lakehouse: Consumes event streams via PySpark 4.1.1 and flushes them to Google Cloud Storage (GCS) as Parquet files using Hive partitioning to minimize downstream scan costs.
• Zero-Copy BigQuery DWH: Integrates BigQuery External Tables to automatically discover GCS partitions, enabling analytics without data duplication.
• Analytics Engineering & BI: Implements staging and metrics mart layers in dbt (calculating running averages of exhaust temperature margins) and visualizes engine degradation in a Streamlit dashboard.

A modern, data-driven static website for a children's theatre studio in Saint Petersburg.

Live Site - studio

Live Site - play

20+ Docker services across 3 bare-metal Ubuntu nodes. Custom domain, automatic TLS, remote access from anywhere via sing-box tunnel.

Three physical machines. All running Ubuntu 24.04 LTS, operating as a distributed cluster. Every service is containerized, reverse-proxied through Traefik with automatic Let's Encrypt TLS, and accessible from anywhere through the https://... domain.

• Cloud & Productivity Self-Hosting: Nextcloud. Vaultwarden. Mattermost.
• AI Agent Infrastructure: OpenClaw. Hermes.
• Media Server Suite: Immich . Navidrome. Audiobookshelf.
• Network & Security: sing-box.
• Observability Stack: Grafana dashboards visualize system metrics collected by Prometheus and node_exporter on each machine.
• Privacy Research Tools: SearXNG.
• Power Resilience: A CBR 850 UPS monitored via NUT (Network UPS Tools). Activates custom bash script for the graceful cluster poweroff.
• Apache Spark 4.1.1 cluster.

Feel free to reach out here:

• 💼 LinkedIn: @stan-buren
• 💬 Telegram: @stan_buren
• 🐦 X (Twitter): @stan_buren
• 📧 Email: stan-buren-dev@proton.me