I build data and AI systems for real business problems, and I measure whether they work:
baselines, held-out tests, and an honest account of what didn't.
| Project | What it does | Result |
|---|---|---|
| it-service-desk-triage | ITIL ticket triage API: an LLM judges category, impact and urgency; fixed rules set priority, resolver group and SLA deadlines. English and German tickets. | 0 of 16 P1/P2 incidents under-called (keyword baseline: 12) |
| policy-compliance-rag | Q&A over company policies, GDPR, the AI Act and NIS2. Every answer cites document, page and section; questions the policies don't cover are refused. | 25/25 answers correct and cited, 8/8 uncovered questions refused |
| resume-screener-eu-ai-act | Ranks CVs against a job description but never decides: anonymisation before the LLM call, evidence-checked claims, audit log, built for the EU AI Act's high-risk rules. | 53 tests · CI |
| it-ops-compliance-assistant | LangGraph agent for IT operations and ISO 27001, using role-restricted MCP tools, with audit logging and human-approved ticket drafts. | 43 tests · CI |
| llm-agent | Tool-calling agent (LangChain + Groq) that chooses between web search, Wikipedia and a calculator to answer questions. | Handles failing tools gracefully |
| n8n-llm-automation | Low-code workflow: support requests arrive by webhook, an LLM classifies and summarises them, urgent ones are routed to an external API. | No-code / low-code automation |
| Project | What it does | Result |
|---|---|---|
| db-delay-tracker | Deutsche Bahn punctuality pipeline for Nürnberg, Fürth, Erlangen and Bamberg: DuckDB extract from a national archive, dbt models with data tests, monthly scheduled refresh. | 1.57M train stops · 20 dbt tests |
| job-skill-tracker | Daily pipeline that collects German job postings and extracts skills, seniority and language requirements, with an auto-updated report. | Runs daily · 47 tests |
| energy-market-airflow-etl | Airflow pipeline for the German electricity market (SMARD): idempotent daily loads, data-quality checks, cheapest hours for energy-intensive production. | Load shifting would have cut a 2 MW job's October energy cost by 35 % |
| kg-etl-pipeline | ETL from JSON, YAML and XML into an RDF knowledge graph, with Pydantic validation and a FastAPI interface. | 13 tests · CI |
| Project | What it does | Result |
|---|---|---|
| credit-default-model | Default risk for 30,000 credit card customers: baselines, tuned LightGBM, calibration, a threshold set by business cost, SHAP, MLflow and a model card. | AUC 0.784 · cost 17 % below the analyst rule |
| demand-forecast | 28-day sales forecasts for 3,049 Walmart products (M5), backtested on three earlier windows against simple baselines. | Store-level error 6.8 % vs 15.1 % for a 28-day average |
| german-news-classifier | Fine-tuned German BERT vs a TF-IDF baseline on 10,000 news articles, compared on accuracy, speed, size and error types. | 90.7 % vs 87.1 % accuracy, at 620× the CPU cost |
| dax-risk-monitor | Weekly 99 % Value-at-Risk for 15 DAX stocks: historical simulation, EWMA, GARCH and filtered historical simulation, backtested over 20 years with Kupiec, Christoffersen and Basel traffic-light tests. Refreshes every week. | Only filtered HS passes: 15 exceptions vs 10.4 expected |
| industry4-semiconductor-quality-monitoring | Early warning for failing wafers on real fab data (UCI SECOM): PCA monitoring, autoencoder and Isolation Forest, with sensor root-cause hints. | ~2× more failures caught than random inspection |
| f1-race-predictor | Predicts Formula 1 finishing positions from qualifying, championship standings, recent form and track history (FastF1 data), with Random Forest and XGBoost. | Matches the strong "finish = grid" baseline (Spearman 0.63 vs 0.62) |
| FLM | Federated learning on non-IID data: a centralised PyTorch model vs synchronous (FedProx) and asynchronous (FedAsync) training across 6 clients. | Centralised vs federated comparison |
| Project | What it does | Result |
|---|---|---|
| retail-sales-powerbi | Power BI sales dashboard on 1 million real retail transactions: cleaned star schema, DAX measures, cohort retention. | Revenue +2.5 % driven by order size, not order count |
| controlling-report | Monthly controlling report from SAP cost exports: reconciliation, data checks, budget vs actual vs prior year, red/amber/green status and written commentary in Excel. | Every file reconciled · 19 tests |
| IT-asset-compliance-tracker | Excel IT asset register with Python automation: ISO 27001 evidence packs, end-of-life checks and licence reconciliation. | Audit-ready Excel evidence pack |
| email-ab-test | Analysis of a randomised e-mail experiment with 64,000 customers: significance tests, multiple-testing correction, power analysis and a decision memo. | +$0.77 revenue per customer vs no e-mail |
| Project | What it does | Result |
|---|---|---|
| order-automation-case-study | PRD for an order-entry AI agent: users, success metrics, phased rollout, risks (GDPR, works council) and an Excel business case. | €89k a year, 8-month payback |
| sample-tracker | Excel tracker for product development samples with a dashboard and data checks, plus an auto-generated weekly PowerPoint status deck. | Weekly deck in one command |
| IT-governance-status-reporter | Turns Jira tickets into PMO reports: KPI trends, RAG status and a one-page steering committee PDF. | Jira Cloud or offline demo |
| ai-use-case-assessment-eu-ai-act | Prioritises AI use cases by value and feasibility, then checks each against EU AI Act risk tiers and GDPR. | Portfolio matrix and steering one-pager |
| Project | What it does | Result |
|---|---|---|
| qa-coverage-tool | Generates test scenarios from feature descriptions with an LLM, tracks coverage gaps, and runs Python and TypeScript regression suites. | pytest + Jest in CI |
| playwright-e2e-suite | End-to-end browser tests for a web app with Playwright and pytest: creating, editing, filtering and deleting items. | Runs in GitHub Actions |
| Domain | Technologies |
|---|---|
| 🤖 AI & LLMs | LangGraph · LangChain · MCP · RAG · Hugging Face Transformers · LLM evaluation · Azure OpenAI · Groq |
| 🏗️ Data Engineering | SQL · dbt · DuckDB · Apache Airflow · Parquet · ETL/ELT · Data quality tests · GitHub Actions |
| 📈 Machine Learning | scikit-learn · LightGBM · XGBoost · PyTorch · MLflow · SHAP · Time series · NLP · GARCH / Value-at-Risk |
| 📊 Analytics & BI | Power BI · DAX · Excel · pandas · A/B testing · Statistics |
| 🗂️ Product & Project | PRDs · Business cases · Jira · Agile/Scrum · Stakeholder reporting · EU AI Act · GDPR |
| ⚙️ Engineering | FastAPI · Pydantic · pytest · Docker · REST APIs · CI/CD |
🎓 Education
- Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) | Nürnberg, Germany
- M.Sc. in International Information Systems (Oct 2024 – Present)
- Focus: Machine Learning, Process Mining & Analytics, Federated Learning, Enterprise Knowledge Management.
💼 Industry Experience
- Cognizant | Programmer Analyst (Prompt Engineer) |
Apr 2023 – Sep 2024(Pune, India)- Built context-aware enterprise chatbots with LangChain and Azure OpenAI, cutting response latency by 40% through prompt optimization and workflow automation.
- Deployed GenAI solutions on Azure Cloud with function calling, real-time retrieval and risk-control guardrails.
- Designed and tested Python AI/ML pipelines across the Agile SDLC.
- Cognizant | Data Integration Intern |
Mar 2022 – Sep 2022(Pune, India)- Developed automated data ingestion and validation pipelines with Python, SQL and shell scripting for large-scale data processing.
Open to Working Student, Internship and Master's Thesis positions in AI Engineering, Data Engineering, Data Science, BI and Product / Project Management (Germany / EU / Remote).
