Works

A Sample of My Previous Works

Senior-Friendly Web Content Transformation System

internet browsing with Senior tablet — Demonstration of a simplified web browsing interface designed for seniors.

Senior-Friendly Web Content Transformation System

June 2023 - January 2025

Company Overview

Lylu GmbH develops senior-friendly software that redesigns existing internet content to make digital technology more accessible for older adults, thereby improving their quality of life.

Role: Machine Learning Engineer and Data Scientist

Solely responsible for developing and implementing an innovative software tool that transforms web pages into a specially adapted, barrier-free user interface tailored to the needs of older people.

Key Technologies

PyTorch
LangChain
Google Cloud Platform (GCP)
Microsoft Azure (Azure AI Services)
Natural Language Processing (NLP)
Transformers
Vector Databases
APIs

Key Innovations

Advanced NLP Techniques: Implemented text embedding, similarity search, and semantic similarity for improved content understanding and transformation.
Model Fine-Tuning: Worked on fine-tuning open-source models to better suit the specific needs of senior-friendly content adaptation.
Contextual Understanding AI: Developed methods to autonomously transform web pages into a barrier-free interface specifically designed for older adults.
Efficient Data Architecture: Utilized vector databases to enhance platform functionality, enabling complex web content to be converted into a simplified, structured JSON format.

Implementation Highlights

Designed and implemented a comprehensive machine learning pipeline for web content analysis and transformation.
Developed APIs to seamlessly integrate the transformation tool with existing web infrastructures.
Created a scalable system capable of processing and adapting diverse web content types.
Implemented robust error handling and performance optimization techniques to ensure reliable operation under various conditions.

Key Achievements

Successfully developed and implemented the core technology driving Lylu GmbH's senior-friendly web transformation tool.
Consistently exceeded performance targets and project goals, even under challenging circumstances.
Demonstrated high levels of initiative, efficiency, and independent problem-solving throughout the project lifecycle.
Received recognition for delivering outstanding work results and achieving project objectives.

Skills Demonstrated

Advanced proficiency in machine learning and data science techniques
Expertise in NLP and content transformation technologies
Strong ability to conceptualize and implement complex AI systems
Proficiency in working with and fine-tuning large language models
Experience with cloud platforms (GCP) for deploying ML solutions
Ability to integrate multiple cutting-edge technologies into a cohesive system
Excellent analytical and problem-solving skills
Strong communication and teamwork abilities

AI Engineer: Specialist in Advanced LLM Applications and Retrieval Systems

January 2023 - Present (Freelance)

Role Overview

As a freelance AI Engineer, I specialize in developing cutting-edge AI systems leveraging state-of-the-art Large Language Models (LLMs), advanced retrieval techniques, and custom-tailored solutions for complex information processing tasks. My work focuses on pushing the boundaries of what's possible with AI, creating systems that are not only intelligent but also efficient, scalable, and adaptable to real-world business needs.

Key Technologies and Frameworks

Large Language Models: GPT-4, Claude, LLaMA 2, Mixtral 8x7B
APIs: OpenAI API, Anthropic Claude API
Python 3.9+, PyTorch 2.0+
Hugging Face Transformers, LangChain, LlamaIndex
Vector Databases: Pinecone, Weaviate, Milvus
Elasticsearch for hybrid search
Neo4j for graph-based knowledge representation
FastAPI for backend services
Docker and Kubernetes for deployment
MLflow for experiment tracking and model management
Ray for distributed computing

Major Projects

Advanced Hybrid Retrieval System with LLM Integration

Objective: Develop a state-of-the-art retrieval system combining dense and sparse retrieval methods with LLM-powered reranking.

Implemented a hybrid search architecture using dense vector embeddings (via SentenceTransformers) and sparse BM25 retrieval.
Developed a custom reranking module using the Mixtral 8x7B model, fine-tuned on domain-specific data for improved relevance assessment.
Integrated a novel top-K retrieval optimization algorithm, dynamically adjusting K based on query complexity and result diversity.
Implemented efficient caching mechanisms and query planning to reduce latency in high-throughput scenarios.
Result: Achieved a 40% improvement in Mean Reciprocal Rank (MRR) compared to traditional retrieval methods, with sub-100ms latency for most queries.

Multi-Modal AI Assistant with LLaMA 2 and Claude Integration

Objective: Create an AI assistant capable of processing and generating multi-modal content, leveraging the strengths of different LLMs.

Fine-tuned LLaMA 2 (70B parameter version) on a custom dataset for domain-specific knowledge and task handling.
Integrated Claude API for advanced reasoning tasks and handling of complex, nuanced queries.
Developed a novel prompt routing system that dynamically selects the most appropriate LLM based on query type and complexity.
Implemented CLIP for image understanding and a custom GPT-4 Vision pipeline for image-to-text tasks.
Created a sophisticated prompt engineering system with dynamic few-shot learning capabilities.
Result: A versatile AI assistant capable of handling text, image, and mixed-modal inputs with high accuracy and contextual understanding.

Enterprise Knowledge Graph with LLM-Powered Query Interface

Objective: Develop a comprehensive knowledge management system combining graph databases, vector search, and LLM-based natural language understanding.

Designed a scalable knowledge graph using Neo4j, integrating data from various enterprise sources.
Implemented a hybrid retrieval system combining graph traversal algorithms with dense vector similarity search in Pinecone.
Developed a custom embedding pipeline using a fine-tuned BERT model for domain-specific entity and relationship embedding.
Created an LLM-powered natural language interface using GPT-4, allowing complex query formulation and result explanation.
Implemented an innovative "graph-of-thoughts" reasoning approach, enabling the LLM to perform multi-hop reasoning over the knowledge graph.
Result: Enabled natural language querying of complex enterprise knowledge, reducing time-to-insight by 60% for data analysts and decision-makers.

Technical Innovations

Adaptive Retrieval Fusion: Developed a novel algorithm that dynamically adjusts the weighting of different retrieval methods (dense, sparse, graph-based) based on query characteristics and past performance.
LLM Ensemble Techniques: Created a sophisticated system for combining outputs from multiple LLMs (e.g., GPT-4, Claude, LLaMA 2) using adaptive weighting and confidence scoring.
Efficient Fine-Tuning Pipeline: Implemented a streamlined process for fine-tuning large language models using techniques like LoRA and QLoRA, significantly reducing computational requirements while maintaining performance.
Contextual Prompt Optimization: Developed an AI-driven system that automatically generates and refines LLM prompts based on user context, query intent, and ongoing conversation flow.
Federated Learning for LLMs: Designed a privacy-preserving method for continually improving language models using decentralized data from multiple clients, addressing data sensitivity concerns in enterprise environments.

Skills and Expertise

Deep expertise in Large Language Models, including fine-tuning, prompt engineering, and efficient deployment strategies
Advanced knowledge of information retrieval techniques, including vector search, hybrid retrieval, and re-ranking methodologies
Proficiency in designing and implementing end-to-end AI pipelines, from data preprocessing to model deployment and monitoring
Strong skills in optimization techniques for high-performance AI systems, including distributed computing and efficient resource utilization
Experience with MLOps practices, including experiment tracking, model versioning, and automated deployment in cloud environments
Expertise in graph algorithms and their application in knowledge retrieval and reasoning systems
Familiarity with ethical AI principles, including bias mitigation and fairness in language models

Impact and Results

Consistently delivered projects that exceeded client expectations, with an average satisfaction score of 4.9/5
Reduced query response times by 70% while improving relevance scores by 35% in large-scale retrieval systems
Published a technical paper on "Adaptive Hybrid Retrieval for Enterprise Knowledge Systems" in a top-tier AI conference
Contributed to open-source projects, including performance optimizations for the LlamaIndex library
Mentored 5 junior AI engineers, focusing on advanced LLM techniques and efficient system design

Continuous Learning and Development

Completed advanced courses in "Efficient Fine-tuning of Large Language Models" and "Graph Neural Networks for Knowledge Representation"
Regular participant and speaker at AI conferences, including NeurIPS, ICLR, and the LangChain Conference
Active member of AI research communities, contributing to discussions on LLM advancements and ethical AI development

Advanced NLP: Optimizing Machine Learning Models for Simple Language Classification

September 2022

Project Overview

This research project focused on developing and optimizing machine learning models for classifying and processing simple language in news articles. The study aimed to enhance NLP techniques specifically for simple language texts, with implications for accessibility and information retrieval.

Key Technologies and Libraries

Python 3.8
pandas 1.3.5 for data manipulation
scikit-learn 1.0.2 for machine learning models and evaluation
spaCy 3.2.0 for advanced NLP tasks
NLTK 3.6.5 for text preprocessing
TensorFlow 2.8.0 for deep learning models
Gensim 4.1.2 for word embeddings

Machine Learning Models and Techniques

Traditional ML Models:
- Logistic Regression with L2 regularization
- Random Forest Classifier (n_estimators=100, max_depth=10)
- Support Vector Machine (kernel='rbf', C=1.0)
- Gradient Boosting Classifier (n_estimators=100, learning_rate=0.1)
Deep Learning Models:
- Convolutional Neural Network (CNN) for text classification
- Long Short-Term Memory (LSTM) network
- Bidirectional LSTM with attention mechanism
Word Embeddings: Utilized pre-trained GloVe embeddings and trained custom Word2Vec models on the simple language corpus

Feature Engineering and Preprocessing

TF-IDF vectorization with n-gram range (1,3)
Custom feature extraction for simple language characteristics:
- Sentence complexity scores
- Readability metrics (Flesch-Kincaid, SMOG)
- Part-of-speech tag distributions
- Named Entity Recognition (NER) densities
Text normalization: lowercasing, punctuation removal, lemmatization
Stop words removal with custom list for simple language

Experimentation and Optimization

Implemented k-fold cross-validation (k=5) for robust model evaluation
Performed hyperparameter tuning using GridSearchCV and RandomizedSearchCV
Developed a custom metric for simple language classification accuracy
Employed ensemble methods, including Voting Classifier and Stacking, to improve overall performance
Conducted ablation studies to identify most influential features for simple language classification

Key Findings and Results

Bidirectional LSTM with attention mechanism achieved the highest accuracy (92.3%) on simple language classification
Custom Word2Vec embeddings trained on simple language corpus outperformed pre-trained GloVe embeddings by 3.7% in classification tasks
Ensemble of SVM, Random Forest, and BiLSTM improved overall F1-score by 2.1% compared to best individual model
Readability metrics and custom simple language features improved model performance by 5.2% when added to traditional TF-IDF features

Impact and Practical Applications

Accessible Content Recommendation: Developed a prototype recommendation system using the top-performing model to suggest simple language articles to users
Automated Simplicity Checker: Created a tool that uses the trained models to assess the simplicity level of given texts and suggest improvements
Cross-lingual Simple Language Detection: Extended the models to identify simple language content across multiple languages using multilingual word embeddings

Advanced Skills Demonstrated

Design and implementation of complex machine learning pipelines for NLP tasks
Proficiency in feature engineering for specialized language processing
Advanced model selection, evaluation, and ensemble techniques
Deep learning architecture design and optimization for text classification
Development of custom evaluation metrics for specific NLP challenges
Application of machine learning models to real-world accessibility problems

Optimizing Classification Performance Through Advanced Synthetic Data Generation Techniques

June 2022

Project Overview

Conducted a comprehensive study to evaluate and compare various synthetic data generation methods for enhancing classifier performance in natural language processing tasks. This research provides a systematic framework for selecting optimal data augmentation techniques, significantly reducing the need for trial-and-error approaches in machine learning pipelines.

Key Technologies and Libraries

Python 3.8+
scikit-learn 0.24.2 for machine learning models and evaluation metrics
NLTK 3.6.2 for natural language processing tasks
gensim 4.0.1 for word embeddings and linguistic transformations
transformers 4.6.1 for state-of-the-art NLP models
googletrans 3.1.0a0 for back translation
matplotlib and seaborn for data visualization
scikit-learn-extra 0.2.0 for t-SNE implementation

Datasets and Preprocessing

Utilized three binary classification datasets with varying sample sizes and class distributions
Implemented robust preprocessing pipeline:
- Text cleaning (removing special characters, lowercasing)
- Tokenization using NLTK's word_tokenize
- Stop word removal
- Lemmatization using WordNetLemmatizer
Performed exploratory data analysis to understand class imbalances and text characteristics

Advanced Data Augmentation Techniques

AEDA (Add Every Distinct Augmentation):
- Implemented custom algorithm for random punctuation insertion
- Controlled insertion rate to maintain text coherence
EDA (Easy Data Augmentation):
- Developed modular functions for synonym replacement, random insertion, random swap, and random deletion
- Implemented adaptive augmentation rate based on sentence length
WordNet-based Augmentation:
- Utilized NLTK's WordNet interface for synonym retrieval
- Implemented part-of-speech aware synonym replacement
- Developed word sense disambiguation mechanism to ensure contextually appropriate synonyms
Back Translation:
- Integrated Google Translate API for multi-language translation chains
- Implemented error handling and rate limiting to manage API requests
- Experimented with various intermediate languages to maximize diversity

Machine Learning Models and Evaluation

Feature Extraction:
- TF-IDF vectorization with n-gram range (1,2)
- Word embeddings using Word2Vec trained on the augmented corpus
Classification Models:
- Logistic Regression with L2 regularization
- Support Vector Machine (SVM) with RBF kernel
- Random Forest Classifier
- Gradient Boosting Classifier
Evaluation Metrics: Accuracy, Precision, Recall, F1-score, ROC-AUC
Cross-validation: Stratified 5-fold cross-validation to ensure robust performance estimation

Advanced Visualization and Analysis

t-SNE Visualization:
- Implemented t-SNE algorithm for dimensionality reduction of high-dimensional text features
- Optimized perplexity and learning rate parameters for each dataset
- Generated 2D and 3D visualizations to analyze class separability
Performance Analysis:
- Developed custom scripts to aggregate results across multiple runs and augmentation methods
- Created comparative visualizations (box plots, heatmaps) to illustrate the impact of each augmentation technique
- Conducted statistical significance tests (paired t-tests) to validate improvements

Key Findings and Results

Back Translation consistently outperformed other methods, improving F1-scores by an average of 7.3% across all datasets
WordNet-based augmentation showed significant improvements for datasets with limited vocabulary diversity
AEDA proved effective for datasets with formal language, improving precision by up to 5.2%
EDA methods demonstrated balanced improvements across all metrics, with an average increase of 4.1% in overall accuracy
t-SNE visualizations revealed improved class separability for all augmentation methods, with Back Translation showing the most distinct clusters

Impact and Practical Applications

Developed a comprehensive guide for selecting optimal data augmentation techniques based on dataset characteristics
Created a modular Python package for easy integration of augmentation methods into existing NLP pipelines
Demonstrated the potential for significant performance improvements in low-resource NLP scenarios
Findings applicable to various domains including sentiment analysis, content moderation, and document classification

Future Work

Exploration of more advanced augmentation techniques using generative models (e.g., GPT-3 for text generation)
Investigation of the impact of augmentation on model robustness and generalization to out-of-distribution samples
Development of adaptive augmentation strategies that dynamically select methods based on input characteristics
Extension of the study to multi-class and multi-label classification tasks

Advanced Skills Demonstrated

Expertise in natural language processing and text augmentation techniques
Proficiency in designing and implementing comprehensive machine learning experiments
Advanced data visualization and analysis skills, including dimensionality reduction techniques
Strong statistical analysis capabilities for validating experimental results
Experience in developing modular, reusable code for complex NLP tasks
Ability to synthesize findings into actionable insights for practical applications

Software Developer Intern at KPMG AG Wirtschaftsprüfungsgesellschaft

January 2020 - December 2021

Company Overview

KPMG is a global network of professional firms providing Audit, Tax, and Advisory services. With over 227,000 employees across 146 countries, KPMG is one of the Big Four accounting organizations. In Germany, KPMG is a leading firm with more than 12,500 employees across 26 locations.

Role: Software Developer

Worked in the Financial Services, Tax Asset Management department at the Frankfurt am Main office. This department specializes in tax advisory services for the financial services sector, including banks, insurance companies, asset management firms, and real estate companies.

Key Responsibilities

Development of new software applications using C#
Enhancement and maintenance of existing C# software solutions
Implementation of database connectivity using Entity Framework
Collaboration with cross-functional teams to understand and implement business requirements
Participation in the full software development lifecycle, from requirement analysis to deployment

Technical Skills Utilized

C# programming language
Microsoft .NET Framework
Entity Framework for ORM (Object-Relational Mapping)
SQL Server database management
Visual Studio IDE
Version control systems (e.g., Git)
Agile development methodologies

Key Projects

Tax Calculation Engine Enhancement: Contributed to improving the performance and accuracy of a C#-based tax calculation engine used for financial instrument analysis.
Database Optimization: Implemented efficient database queries and optimized Entity Framework usage, resulting in a 30% improvement in data retrieval times.
Reporting Tool Development: Assisted in creating a new reporting tool for asset management clients, integrating various data sources and providing customizable report generation capabilities.

Key Achievements

Demonstrated strong analytical and problem-solving skills in developing software solutions for complex financial scenarios.
Successfully integrated into the team environment, collaborating effectively with both technical and non-technical colleagues.
Received positive feedback for the quality and reliability of work delivered, consistently meeting project deadlines.
Showed initiative in learning about the financial services sector and its specific technological needs.

Professional Skills Demonstrated

Strong programming skills with a focus on C# and .NET technologies
Ability to quickly adapt to new technologies and business domains
Excellent problem-solving and analytical thinking capabilities
Effective communication skills in a professional, multinational environment
Attention to detail and commitment to producing high-quality work
Ability to work independently and as part of a team
Time management and ability to meet deadlines in a fast-paced environment

Impact and Learning

This internship provided valuable exposure to the intersection of technology and financial services. It enhanced my understanding of how software solutions can address complex business needs in the financial sector. The experience at KPMG has significantly contributed to my professional growth, improving both my technical skills and my ability to work in a corporate environment.

KrakenBot: Advanced Cryptocurrency Trading System with Real-Time Market Analysis

April 2021 - Present

Project Overview

Developed a comprehensive, production-ready cryptocurrency trading bot that interfaces directly with the Kraken exchange API. This system combines real-time market data analysis, advanced trading strategies, and machine learning predictions to execute automated trades and provide insightful market analytics.

Key Technologies and Libraries

Python 3.9+
FastAPI for high-performance API development
SQLAlchemy for database ORM
PostgreSQL for robust data storage
Pydantic for data validation and settings management
Alembic for database migrations
TA-Lib for technical analysis indicators
TensorFlow 2.x for machine learning models
Docker and Docker Compose for containerization
GitHub Actions for CI/CD

System Architecture

Modular Design: Separated concerns into distinct components (API, database, trading logic, ML predictions)
Real-time Data Processing: Implemented websocket connections for live market data streaming
Scalable Database Schema: Designed efficient models for storing order book data, trades, and market trends
RESTful API: Created endpoints for bot control, data retrieval, and strategy management
Containerized Deployment: Utilized Docker for consistent development and easy deployment

Advanced Trading Strategies

Market Making: Implemented a sophisticated market making strategy with dynamic spread adjustments
Technical Analysis: Incorporated various TA indicators (e.g., RSI, MACD, Bollinger Bands) for trend identification
Order Book Analysis: Developed algorithms to analyze order book depth and liquidity
Machine Learning Integration: Used LSTM networks for short-term price movement predictions
Risk Management: Implemented position sizing and stop-loss mechanisms to control risk exposure

Key Features and Capabilities

Multi-Asset Support: Capable of trading multiple cryptocurrency pairs simultaneously
Real-time Performance Monitoring: Dashboard for live tracking of bot performance and market conditions
Backtesting Engine: Allows for strategy testing on historical data before live deployment
Automated Trade Execution: Executes trades based on predefined strategies and market conditions
Dynamic Strategy Adjustment: Adapts strategies based on changing market volatility and trends
Detailed Logging and Reporting: Comprehensive logs for auditing and performance analysis

Machine Learning Model

Architecture: LSTM neural network for time series prediction
Features:
- Historical price data (OHLCV)
- Technical indicators (RSI, MACD, Bollinger Bands)
- Order book imbalance
- Volume profile
Training Process: Continuous retraining on recent market data to adapt to changing conditions
Integration: Predictions used as additional input for trading decision logic

Notable Achievements and Results

Successfully deployed and operated on the Kraken exchange, handling real-time trading of major cryptocurrency pairs
Achieved consistent profitability over a 3-month testing period, outperforming buy-and-hold strategy by 12%
Developed a robust system capable of handling high-frequency trading with low latency (avg. execution time < 100ms)
Implemented advanced risk management, resulting in a maximum drawdown of only 5% during volatile market conditions
Created a flexible framework allowing easy integration of new trading strategies and indicators

Challenges Overcome

High-Frequency Data Handling: Optimized database schema and implemented efficient data processing pipelines to handle large volumes of real-time market data
API Rate Limiting: Developed intelligent request queuing and rate limiting system to comply with exchange API restrictions
Market Volatility: Implemented adaptive algorithms that adjust trading parameters based on detected market regimes
System Reliability: Designed comprehensive error handling and automatic recovery mechanisms to ensure 24/7 operation

Future Enhancements

Integration of natural language processing for sentiment analysis of news and social media
Implementation of reinforcement learning for dynamic strategy optimization
Expansion to support multiple cryptocurrency exchanges for cross-exchange arbitrage
Development of a web-based user interface for easier bot configuration and monitoring

Advanced Skills Demonstrated

Design and implementation of production-grade, high-frequency trading systems
Proficiency in financial markets analysis and algorithmic trading strategies
Advanced Python development with focus on performance optimization
Experience with real-time data processing and API integration
Containerization and microservices architecture design
Machine learning model development and integration in financial applications
Robust database design and optimization for high-volume data

Java Developer and Test Engineer at Mobina

June 2011 - May 2016 (5 years)

Company Overview

Mobina is a local software development company specializing in creating custom solutions for small to medium-sized businesses. The company focuses on developing management information systems, inventory control applications, and basic e-commerce platforms for various industry sectors.

Role: Java Developer and Test Engineer

I joined Mobina as a junior Java developer and grew into a more senior role over my four-year tenure. My responsibilities expanded from basic coding tasks to designing and implementing significant portions of our software solutions, as well as taking on testing and quality assurance duties.

Career Progression

Junior Java Developer (2011-2012): Started with basic coding tasks and bug fixes under close supervision.
Java Developer (2012-2013): Took on more complex programming assignments and began contributing to project planning.
Senior Java Developer (2013-2015): Led small development teams and took responsibility for critical software components.
Java Developer and Test Engineer (2015-2016): Added testing and quality assurance to my developer responsibilities.

Key Responsibilities

Developed Java-based applications for inventory management and e-commerce platforms
Implemented and maintained database systems using MySQL
Created unit tests and performed manual testing to ensure software quality
Collaborated with team members to design and implement new features
Provided technical support and bug fixes for existing applications
Participated in client meetings to gather requirements and present solutions
Mentored junior developers as I gained experience

Technical Skills Developed

Java SE
Java Swing for desktop applications
Basic JSP and Servlets for web applications
MySQL database design and optimization
JDBC for database connectivity
JUnit for unit testing
Apache Tomcat web server
Version control with SVN
Eclipse IDE

Key Projects

Inventory Management System: Contributed to the development of a Java Swing-based desktop application for small retail businesses. Implemented features for stock tracking and report generation.
E-commerce Platform: Assisted in creating a basic online shopping system using JSP and Servlets, integrating payment gateway APIs and managing product catalogs.
Testing Framework Development: Developed a simple automated testing framework using JUnit to improve the efficiency of our quality assurance process.

Key Achievements

Grew from a junior developer to a trusted team member capable of handling complex tasks independently.
Implemented a more structured testing process, reducing post-release bugs by approximately 30%.
Received a "Best Team Player" award in 2013 for my contributions to improving team collaboration and knowledge sharing.
Successfully delivered a critical project under a tight deadline, earning commendation from company management.

Professional Skills Developed

Proficiency in Java development for both desktop and basic web applications
Understanding of software development lifecycle and basic project management
Ability to translate client requirements into technical solutions
Problem-solving skills, particularly in debugging and optimizing code
Effective communication with team members and occasionally with clients
Time management and ability to work on multiple projects simultaneously
Basic mentoring and knowledge sharing with junior team members

Impact and Growth

My time at Mobina marked the beginning of my professional journey in software development. Over four years, I evolved from a novice programmer to a competent Java developer with a growing understanding of software testing. This experience laid a solid foundation in core Java programming, database management, and the basics of web application development. It also helped me develop crucial soft skills such as teamwork, communication, and problem-solving, which have been invaluable in my subsequent career growth.

↑