I'm a PhD-trained data scientist and Google Cloud Certified Professional Machine Learning Engineer with 6+ years of experience building end-to-end data and ML systems in production.
I specialize in taking models beyond notebooks—designing scalable data pipelines, deploying time-series and recommendation systems, and building internal data products that drive measurable business impact. From forecasting revenue, automating workflows to improving recommendation accuracy, I focus on reliable, maintainable, and value-generating ML solutions.
I'm deeply passionate about building robust, scalable technology and applying machine learning where it creates meaningful, measurable impact in real-world products.
I built and deployed a production-oriented AI agent system for job search workflows. The assistant uses a LangGraph-orchestrated pipeline for intent classification, policy-based tool routing, guarded execution, response synthesis, telemetry, token usage tracking, and message persistence. This project showcases machine learning engineering across LLM workflow design, retrieval-augmented tool use, evaluation, backend systems, frontend UX, and cloud deployment.
Agent architecture: The core system is a tool-augmented assistant built with FastAPI, LangChain, and LangGraph. I implemented multiple AI workflows, including natural-language database querying with safety checks, AI-assisted CSV import for messy job-tracking data, schema-constrained webpage extraction for job posting ingestion, profile-aware drafting using CV and portfolio context, and semantic job search powered by Sentence Transformers embeddings and pgvector.
Evaluation and reliability: I treated the LLM workflow as a production ML system, adding deterministic policy routing, required-tool guard nodes, retry logic for transient provider failures, structured node/tool telemetry, trace propagation, and token usage tracking for cost visibility. I also built an offline routing evaluation harness with labeled datasets, saved experiment runs, route accuracy metrics, critical misroute tracking, clarification-rate analysis, and model comparison across LLM providers.
Cloud deployment: The application is containerized with Docker and deployed on AWS using ECS Fargate for frontend/backend services and Amazon RDS PostgreSQL for persistence. Delivery is automated via GitHub Actions with OIDC-based AWS auth, CloudFormation infrastructure-as-code, ECR image builds, CodeDeploy blue/green deployments, Secrets Manager for secret management, and CloudWatch for centralized logging.
I built an end-to-end machine learning application for forecasting day-ahead electricity load in the Germany-Luxembourg (DE-LU) bidding zone. The app ingests load data from the ENTSO-E Transparency Platform and weather data from Open-Meteo, engineers time-series features, and trains forecasting models for operational use. It includes a FastAPI backend for serving forecasts and metrics and a Streamlit dashboard for interactive monitoring and exploration.
