Projects

The Challenge: Slow, Manual Risk Detection in a Critical Process

In German mortgage lending, the Grundbuchauszug (land registry extract) is a critical legal document. It contains the official history of a property, including ownership, rights, and most importantly any existing liens or encumbrances that could pose a risk to the bank.

The traditional workflow involved credit department specialists manually downloading these multi-page, often scanned, PDF documents for each loan application. They would then spend valuable time reading through them to find a specific list of "red flag" keywords (e.g., notices of insolvency, compulsory auctions). This process was:

• Slow: Taking hours or even days, delaying the entire loan approval process.
• Error-Prone: A critical keyword missed by human review could lead to significant financial risk.
• Inefficient: Specialists were spending time on a repetitive, low-value task instead of complex case analysis.

The goal was to create a system that could automate this entire process, delivering critical risk alerts to the credit department the moment a document enters the bank's system.

The Solution: An End-to-End Automated Analysis Pipeline

I designed and built a fully automated, serverless solution that handles the entire workflow from document ingestion to final risk alert, requiring zero manual intervention. The process unfolds in several intelligent steps:

1. Automated Document Ingestion: The system continuously monitors third-party lending platforms for new loan applications. When a new Grundbuchauszug is submitted, it is automatically downloaded.
2. Cloud-Based Text Extraction: Each downloaded PDF is sent to a cloud-based OCR (Optical Character Recognition) service. The service analyzes the document, even if it's a low-quality scan, and returns a structured JSON file containing all the extracted text, lines, and their page locations.
3. Intelligent Keyword Search: The core of the system is an intelligent search algorithm. Instead of naively searching the entire document, the code first identifies the most critical sections, Section II and III of the document, where encumbrances and liens are listed. The keyword search is then focused exclusively on these high-value pages, dramatically increasing both speed and accuracy.
4. Metadata Extraction: To ensure every finding is traceable, the system also identifies and extracts the unique land registry sheet number from the document, allowing multiple extracts within a single application to be distinguished.

The Technology Stack & Implementation in AWS

• Compute: AWS Lambda was used for all business logic. The process was split into two distinct functions: a Fetcher Lambda, triggered by a scheduled Amazon EventBridge rule every hour to check for new documents; and a Processor Lambda, automatically triggered by an Amazon S3 event the moment the Fetcher uploads a new document.
• Storage: Amazon S3 acts as the central hub. It stores the incoming PDF documents, the keywords.csv configuration file, and the final JSON analysis results.
• Security: All sensitive credentials (API keys, passwords) are securely stored and managed in AWS Secrets Manager, completely separating them from the application code.
• Dependencies: All Python libraries, including custom internal modules and complex packages (pandas, cryptography), are managed as AWS Lambda Layers, built in a Linux-based cloud environment (AWS CloudShell) for runtime compatibility.

The Impact: From Days to Minutes

• Real-Time Alerts: The credit department can receive an automated information with any "red flag" findings, including the specific keyword found and the page number, within an hour of the document arriving at the bank.
• Increased Efficiency: The system completely eliminated the manual review time for this task, freeing up specialists to focus on higher-value credit analysis.
• Reduced Risk: By automating the search, the risk of human error in missing a critical term was reduced to zero, strengthening the bank's risk management posture.

Less

This project demonstrates a full-cycle data analysis, from initial statistical validation and deep-dive investigation to predictive modeling and strategic business recommendation. The goal was to understand and manage the risk associated with specific loan products governed by the EU's Product Oversight and Governance (POG) directive.

Phase 1: The Initial Question — Do We Have a Statistically Proven Problem?

The Business Context: The bank identified several loan categories as POG-relevant, meaning they required closer monitoring. These included "Hochausläufer" loans (high loan-to-value), loans with high broker commissions, and loans for older borrowers. The initial business question was simple but crucial: Are these POG loans genuinely riskier than our standard, "Normal" loans?

My Approach: I used a proportions Z-test to compare arrears rates between POG categories and the Normal portfolio.

Null Hypothesis (H₀): No significant difference in arrears rates between a POG category and Normal loans.

Result: For key POG categories like "Hochausläufer," very small p-values (≪ 0.01) allowed rejecting H₀.

Outcome: Statistically irrefutable proof that elevated arrears in POG categories were not random.

Phase 2: Defining the Boundary — The "Simulated Z-Test" Threshold

Challenge: Determine an objective, statistically sound arrears threshold.

Solution: I designed a simulation that redistributes a fixed number of total arrears between POG and Normal groups. After each transfer, a Z-test checks whether the p-value drops below the significance level; the point where it first does defines the defensible threshold.

Outcome: Clear, data-derived red lines per POG category.

Phase 3: The Deep Dive — Understanding the "Why"

Approach: Enriched data and compared profiles of "Hochausläufer" vs Normal loans.

• Younger borrowers are riskier: Under 30 in "Hochausläufer" had ~4.01% arrears (2.5× Normal peers).
• Investment properties drive risk: Non-owner-occupied loans ~4.38% arrears.
• Short-term loans create a vicious cycle: Higher rates → higher payments → pressure on low surplus households.

Outcome: Risk arises from the interaction of product and vulnerable client segments.

Phase 4: Predictive Modeling & Strategic Calibration

Scorecard: Logistic Regression trained on "Hochausläufer" with key drivers (Age, BLA, BUEBERSCH, Eigennutzung, etc.) to score new applications.

Threshold calibration provided three strategies:

• Risk Target: Match approved portfolio arrears to Normal (1.03%).
• Volume Target: Approve a fixed share (e.g., 50%) and observe resulting arrears.
• Profitability Target: Maximize profit by balancing income from good loans vs cost of bad loans.

Outcome: A quantitative menu of options showing Risk–Volume trade-offs for leadership decisions.

Project Impact

Shifted discussion from "this is risky" to a calibrated, quantitative framework. Delivered a working scorecard and calibration analysis to balance risk, volume, and profitability.

Technologies Used: Python, pandas, statsmodels, scikit-learn, Matplotlib, Power BI.

Less

In this project, I developed a Natural Language Processing (NLP) model to automatically classify critical financial documents. The primary goal was to distinguish German land registry extracts (Grundbuchauszüge) from other document types like purchase agreements (Kaufverträge). This automation is vital for streamlining workflows in banking and real estate finance.

The Challenge

Financial institutions process thousands of documents daily. Manually sorting these documents is time-consuming, prone to human error, and creates bottlenecks. This project tackled the challenge by building a machine learning pipeline that can read and understand document content and assign the correct class automatically.

My Approach: From Raw Text to Accurate Predictions

I implemented a classic yet powerful NLP strategy using TF-IDF (Term Frequency–Inverse Document Frequency) combined with a Linear Support Vector Machine (SVC) classifier, built as a robust, reusable pipeline.

1. Data Preparation & Preprocessing
   Text Extraction: parsed JSON from AWS Textract to extract raw text from scanned PDFs.
   Text Cleaning with spaCy: used the German model for lemmatization, removed stop words and punctuation.
2. Feature Engineering with TF-IDF
   Converted cleaned text into numerical vectors. Words like “Grundbuchamt” gain high scores in registry extracts, aiding separation.
3. Model Training & Evaluation
   LinearSVC with 5-fold cross‑validation; evaluated via Accuracy, Precision, Recall, and F1‑Score.

Results & Future Expansion

The initial binary model achieved 100% accuracy on test data for identifying Grundbuchauszüge. A confusion matrix and classification report confirmed perfect separation.

Next: Expand to multi‑class classification (Kaufverträge, Teilungserklärungen, loan applications, salary slips). LinearSVC supports multi‑class via one‑vs‑rest; extend labels accordingly.

Less