Data & AI

Built a full ELT analytics pipeline on GCP to benchmark IT compensation and job demand across Europe. Two sources feed the pipeline: the Stack Overflow 2025 Developer Survey and live job listings scraped from profesia.sk. A layered dbt project on BigQuery (staging → intermediate → marts) handles deduplication, role/seniority classification via custom Jinja macros, incremental loads, and schema tests. Salary benchmarks (median, P25/P75 by country × role × seniority) and job market KPIs are published as a Looker Studio dashboard.

A live, queryable view of European IT salaries and demand — built to understand the market from the inside, using the same tools the market expects.

Long-term big data platform within Commercial Insurance Analytics, built initially for underwriters and gradually extended to serve actuaries, portfolio managers, and country heads. Worked across the full lifecycle: onboarding data sources country by country, setting up cleaning and integration, and evolving a universal data model that linked sources via shared IDs — and via “soft” matching where none existed. Curated datasets were exported to SQL Server and served through reactive RShiny dashboards with access-based visibility; new front-end requests routinely triggered another cycle of onboarding, modelling, and curation. The platform went through two major migrations — from Palantir Foundry to an on-premises Hadoop/Spark stack (Scala, Jenkins), then to Azure Databricks in the cloud — both delivered successfully on the strength of an automated data quality framework.

One universal data model feeding underwriting, actuarial, and portfolio reporting across 10+ countries — stable enough to survive two full platform migrations and keep growing.

Built end-to-end marketing analytics pipelines ingesting event and campaign data from Google Analytics 4 and Google Ads into BigQuery. Applied SQL transformations and scheduled queries to produce clean reporting layers, delivered as Looker Studio dashboards for campaign performance monitoring and client decision-making.

Clients moved from manual spreadsheet exports to automated, always-current dashboards — decisions based on data updated daily, not weekly.

End-to-end predictive system for urban parking, from consulting and feasibility through delivery. Started by scoping two prediction problems with the client's parking operator: lot-level occupancy forecasting (for smarter reservation overbooking) and per-spot departure prediction. Contributed to remediating noisy gate and sensor data — reconstructing hourly occupancy series from session logs and validating cleanliness via masking-based backtesting. Built the models: Facebook Prophet for occupancy (multi-seasonal patterns, Slovak holiday calendar, capacity-normalised demand, 95% confidence intervals) and survival analysis (Kaplan-Meier + Cox Proportional Hazards) for hourly per-spot departure probabilities. Delivered as a containerised FastAPI + MySQL stack with automated per-lot model selection, plus a Streamlit prototype for client demos.

Two production-ready predictive models packaged for direct backend integration — and an interactive prototype that let the client explore predictions before committing to any integration work.

Built this site as a working example of AI-augmented engineering: a Next.js 16 app with Tailwind CSS, deployed on Vercel with auto-deploy from GitHub. A recruiter Q&A bot (in progress) will let visitors ask questions about availability and experience, answered by an LLM grounded in a structured profile JSON.

A portfolio that demonstrates the approach, not just describes it — the bot itself is a live example of practical AI integration.