LLM agent systems for analytics and forecasting.
AI Engineer and Data Scientist specialising in analytics engineering, foundation model forecasting, and LLM agent systems. At InnoGames, my core work centers on end-to-end Power BI data modeling, reporting, and migrating analytics across six games to Microsoft Fabric, augmented by a Claude Code multi-agent toolkit for automated PBIP/TMDL report authoring. In parallel, my company project and M.Sc. thesis at Leuphana University focuses on improving 3-month net revenue forecast accuracy and calibrated uncertainty intervals using frozen foundation models (TabPFN-TS, zero task-specific retraining) and calendar flags, researching whether agent-compiled planning text covariates add measurable increment beyond calendar timing. I bring ten years of analytics experience across gaming, telecommunications, finance, and insurance.
Data modeling, Power BI analytics, and Microsoft Fabric pipelines, extended with LLM agent systems and foundation model forecasting.
At InnoGames, my core work centers on business intelligence and data engineering in Power BI and Microsoft Fabric. I design semantic data models, build comprehensive reporting suites and dashboards for player engagement and monetization across six live games, conduct A/B-test and churn analyses, and lead daily reporting migrations into Microsoft Fabric. To accelerate and govern this workflow, I developed a Claude Code multi-agent toolkit for automated PBIP/TMDL model and report creation, with built-in approval, schema validation, and escalation controls.
Alongside my analytics work, my company project and M.Sc. thesis at Leuphana University focuses on improving revenue forecast accuracy and calibrated uncertainty intervals using frozen foundation models (TabPFN-TS) without task-specific training, combined with calendar flags and researching whether agent-compiled planning text covariates add measurable increment beyond calendar timing.
I am completing an M.Sc. in Management & Data Science at Leuphana University as a Deutschlandstipendium scholar. From April 2027, I am available for full-time AI Engineer and Data Scientist roles in Germany.
Selected projects: LLM agent systems, knowledge graphs, and Bayesian modelling
What I work on
My projects span LLM agent systems, time-series forecasting, Bayesian modelling, and knowledge-graph applications. They include a multi-agent toolkit used by an analytics team, a knowledge-graph question-answering system, a hierarchical Bayesian customer lifetime value model, and a multi-agent assistant for company founders.
Forecasting Free-to-Play Game Revenue with Foundation Models and Agent-Compiled Text Covariates
My company project at InnoGames, also my M.Sc. thesis at Leuphana University, aims to improve 3-month net revenue forecast accuracy and calibrated prediction intervals using a frozen tabular foundation model (TabPFN-TS) with zero task-specific retraining. The research tests whether and how agent-compiled covariates, raw-text features, and embeddings from versioned planning text add measurable value beyond calendar flags, using strict vintage guards and placebo-controlled benchmarks to distinguish genuine document content from calendar timing.
Experiments across 49 monthly origins demonstrate that the frozen foundation forecaster with calendar flags outperforms existing production forecasts across all evaluated games; I am currently writing up the thesis for submission in January 2027.
Fuel price forecasting with LLM event features
Prior Labs TabPFN-3.5 Hackathon: tests whether LLM-extracted event features from a weekly energy-market narrative improve TabPFN-3.5 gasoline price forecasts, with a shuffled-content placebo to separate content from timing.
TabPFN-3.5 vs Bayesian Pareto/NBD
Benchmark for the Prior Labs TabPFN-3.5 Hackathon: when a tabular foundation model beats specialised Bayesian customer models at predicting 90-day customer return, tested on untouched holdout origins with learning curves and bootstrap intervals.
Agentic toolkit for Power BI report changes
Multi-agent Claude Code toolkit used across the analytics team, planning, implementing, and verifying Power BI dashboard changes in PBIP/TMDL without step-by-step human input.
ChefTreff AI Hackathon: Founder Assistant
Multi-agent assistant that helps founders launch a company in Germany, covering business plan, marketing plan, and legal documentation.
Knowledge Graph Question Answering
SPARQL-based question answering over structured knowledge graphs, enabling natural language queries over RDF knowledge bases.
KVG: Lüneburg Public Bus Routing
Applied ML to model and predict local bus routes for KVG, combining public transport geospatial data with predictive modelling.
Neo4j Graph Analysis of Artist Influence Networks
Graph database solution using Neo4j and Cypher to analyse artists' influence and cluster musical lineages.
Hierarchical Bayesian Pareto/NBD: Replication of Abe (2009)
Implemented and validated the hierarchical Bayesian Pareto/NBD model from Abe (2009) on the canonical CDNOW dataset.
ChefTreff AI Hackathon: Product Detection
Built in 24 hours: a computer vision pipeline for identifying broken or damaged objects from product imagery.
Writing on LLM agent systems and applied machine learning
Available in Germany from April 2027, preferably Munich, Hamburg, Düsseldorf, or Berlin.
I am interested in AI Engineer and Data Scientist positions involving LLM-based systems, forecasting automation, or analytics platforms. I combine hands-on AI development with ten years of experience translating business requirements into maintainable data and analytics solutions.
My work fits six kinds of teams well:
- Insurance and reinsurance analytics. Underwriting segmentation at Absolute Insurance and current Bayesian forecasting research, applicable to Munich's insurance cluster and beyond.
- Gaming and consumer products. Revenue forecasting prototypes and analytics infrastructure at InnoGames in Hamburg, including Microsoft Fabric migration and Power BI dashboards.
- AI-native startups and applied AI labs working on LLM systems, agent architectures, and retrieval. I developed and rolled out a multi-agent toolkit now used daily by an analytics team, and my company forecasting project, also my M.Sc. thesis, compares several ways of using planning text in revenue forecasts.
- Telecommunications and network analytics. Eight years at MTS covering market analysis and M&A across the Indian, Serbian, Slovenian, and Moldovan markets, relevant to telecommunications operators and network businesses, including in Munich.
- Global technology firms and platform teams in Munich, Hamburg, Düsseldorf, and Berlin, where forecasting, applied ML, and production analytics scale across products.
- Startups and scale-ups where production ML and agent systems form the core product rather than a pilot.
I work in English at C1 level (IELTS 8.0) and in German at B2 level. My analytics experience spans international teams in finance, telecommunications, insurance, and gaming.