Data professional with 3+ years of industry experience taking ML models from prototype to production, owning the full lifecycle from data ingestion through deployment and monitoring.

🎓 MSc candidate in Artificial Intelligence at the University of Zurich, with a joint master's project at ETH Zurich.

📍 Based in Zurich, open to relocation across the EU and UK.

I work across three tracks: Data Science (predictive modelling, time-series forecasting, statistical evaluation), ML Engineering (RAG systems, model serving, MLOps and observability), and Data Engineering (streaming and batch data pipelines). The projects below are grouped that way.

• london-spatial-house-prices — Spatial ML on 375k real London house sales. Proves that random cross-validation leaks neighbourhood signal (R² 0.70) and measures the honest generalisation gap with spatial CV (R² 0.38), with Moran's I, SHAP, conformal intervals, and residual maps. LightGBM, GeoPandas, PySAL, SHAP.
• timeflow — End-to-end demand forecasting with LightGBM point and quantile models, conformalized prediction intervals, and a newsvendor policy layer that turns forecasts into stocking decisions. LightGBM, conformal prediction, Streamlit.

• scholar-rag — Production RAG over academic literature: hybrid retrieval (BGE-M3 dense + sparse + RRF), cross-encoder reranking, and citation validation. Lifts MRR@10 from 0.925 to 0.967 and abstains correctly on 5/5 out-of-corpus queries, with full Langfuse + Prometheus observability. LangChain, Qdrant, FastAPI, Docker, GitHub Actions.
• job-intelligence-api — ML-powered API that classifies job postings into ML Engineer, Data Engineer, and Data Scientist roles and extracts structured skill tags. FastAPI, MLflow, Docker, CI/CD.

• wiki-stream-pipeline — Real-time streaming pipeline: live Wikipedia edits through Kafka (Redpanda) into Postgres, built for reliability with idempotent delivery, two-tier dead-letter queues, a failure taxonomy, and Dagster-orchestrated data-quality checks. Kafka, Postgres, Dagster, Docker.
• marketpulse-data-platform — Batch data platform with Airflow orchestration, PySpark feature computation, and a Snowflake medallion warehouse (RAW → STAGED → CURATED), with idempotent deduplication and quality gates. Airflow, PySpark, Snowflake, dbt.

Data Science: LightGBM · XGBoost · scikit-learn · PyTorch · SHAP · GeoPandas · PySAL · Prophet · conformal prediction

ML Engineering: LangChain · Qdrant · Hugging Face · FastAPI · MLflow · Langfuse · Prometheus · Docker · GitHub Actions

Data Engineering: Kafka / Redpanda · PySpark · Airflow · Dagster · Snowflake · PostgreSQL · dbt · AWS

Languages: Python · SQL (PostgreSQL, Oracle PL/SQL, T-SQL)

📫 Reach me: LinkedIn · baranozgur.tas@gmail.com