Moritz
Exploring Innovation in Data Analytics & AI Systems
21 | Artificial Intelligence | Data Science | Software Engineering
About
My Introduction
Hooked on AI
I'm deep in the world of AI, exploring its transformative potential across cutting-edge technologies. Beyond AI, I dive into massive software engineering projects that scale complex systems and push architectural boundaries.
Passion for Innovation
I'm passionate about Machine Learning-transforming abstract code into models that recognize images or understand context feels like magic.
Breaking Barriers
I thrive on solving complex problems and turning ideas into tech that makes a difference—those breakthrough moments are what drive me.
Skills
My technical & miscellaneous skillsAI & Data Science
3+ Years XPComputer Vision
92%Natural Language Processing
85%Generative AI
90%Reinforcement Learning
88%Deep Reinforcement Learning
85%RLHF
82%Probability & Statistics
85%Data Analytics & Visualization
80%MLOps
90%Distributed DNN Training
85%Multi Agent Models
85%Neural Networks & HPC
88%Programming & Frameworks
3+ Years XPPython
95%PyTorch
92%TensorFlow
88%NumPy & Pandas
93%R
80%C & C++
90%C#
90%SQL
95%MLflow / Weights & Biases
87%Computing
2+ Years XPDocker
90%Microsoft Azure
80%Amazon Web Services
75%Virtualization
75%Frontend
2+ Years XPHTML
90%CSS
85%JavaScript
75%Angular
85%React Native
85%Backend
2+ Years XPPython - Flask, Fast API
95%Node JS
80%Misc
2+ Years XPGit
90%Linux
90%DevOps
75%Software Development
85%Leadership
95%Communication
90%Time Management
98%Experience
My journey in the academic & professional front
Machine Learning Engineer
Eviden Austria (ATOS)Data Science Engineer
Smart DigitalFreelancer
Paramedic
Research
Academic Research & Thesis WorkFoundational Model for Neurosurgical Video Intelligence
University College London • 2025-2026Foundational Model for Neurosurgical Video Intelligence
Status: In ProgressThis project's goal is to build the first foundational model for neurosurgery. I am training a large model on a unique dataset of over 100 hours of mostly unlabelled surgical video. The model will learn rich, general-purpose visual representations of the complex neurosurgical environment directly from the footage.
Instrument Understanding
This part of the evaluation will test the model's ability to comprehend the tools used in surgery.
- Detection: The model must identify all instruments present in the surgical field.
- Segmentation: It will precisely outline the complete boundary of each tool.
- Pose Estimation: The model will determine the three-dimensional orientation of each instrument.
- Tracking: It will follow the exact path and movement of surgical tools through the video sequence.
Mapping the Operative Environment
This evaluation will assess the model's spatial and anatomical awareness of the surgical site.
- Depth Estimation: The model will reconstruct a 3D view of the operative area from the 2D video feed.
- Tissue Segmentation: It will distinguish between different types of biological tissue.
- Tissue Classification: The model will categorise tissue based on its visual properties.
- Anatomical Recognition: It will identify critical anatomical structures within the brain and surrounding areas.
The final model will provide a powerful base for future neurosurgical AI systems, from robotic assistance to advanced training simulations. I plan to publish the findings in a leading AI journal.
Current Focus: My work now involves preparing the dataset and conducting initial training experiments.
Automatic LLM Personalisation with Human Feedback
UAS Vienna • 2024-2025Automatic LLM Personalisation with Human Feedback
Key Results
PCA plot showing the learned embeddings for the 'lawyer' (user_1) and 'social media analyst' (user_2). Their separation demonstrates the model successfully captured their distinct preferences.
I enhanced a commercial RAG system to address data privacy risks and enable automatic user personalisation. I developed a local LLM with a custom user-embedding layer that learns individual preferences from star ratings. The final system adapted its response style to different user personas, making a lawyer persona 90% more likely to receive a detailed, formal response.
The Problem
The existing system sent user data to third-party APIs, which was unacceptable for clients with sensitive information. The system also had no memory for user preferences between sessions, causing user frustration.
Technical Solution
System Architecture
Architecture diagram showing the RAG system, user feedback loop, local LLM, and custom user-embedding layer integration.
Architecture
My solution uses a Representation Learning Approach. I subclassed a LlamaForCausalLLM to introduce a custom user-embedding layer. This layer concatenates a learned user-specific vector with the standard word embeddings. This method allows a single model to serve all users, which avoids the high cost of training a model for each user.
Training and Optimisation
The model was trained using Reinforcement Learning from Human Feedback (RLHF). I used Group Relative Policy Optimization (GRPO) because its low memory footprint was essential for the single Tesla T4 16GB GPU, as it does not require a reference model like PPO. To make training feasible, I used LoRA, 8-bit precision, and Key-Value (KV) caching, which reduced response times by 1.36 seconds per query.
Results
Example: Persona-Based Responses
Prompt: What is a projectile?
Key Learnings
- Hyperparameter Tuning: RLHF training is very sensitive. A high learning rate caused model collapse, where it only outputted the letter "e".
- Hardware Constraints: Optimisation techniques like GRPO, LoRA, and KV caching were not optional. They were essential to successfully train and run the model on a single 16GB GPU.
- Dataset Selection: The choice of dataset is important. The short answers in the SQuAD dataset were not suitable for training a model on style and verbosity. The QuAC dataset was more effective.
Portfolio
My works, projects & contributions
MarketVision AI - Multi-Modal Trading Agent
Developed a Double DQN trading agent processing a unique 3-channel state by combining candlestick charts, Google Trends data, and news sentiment analysis. Engineered agent with a CNN state encoder and prioritised experience replay, accelerating training and improving trading performance by an average of 50%. Applied Explainable AI (XAI) techniques to interpret model behaviour, using attention maps to visualise candlestick patterns and to quantify the agent's reliance on each input channel. Validated strategy through extensive backtesting and demonstrated its ability to maintain profitability during live paper trading.
View
Smart Digital - Geospatial Object Detection For Railways
Utilizes computer vision and instance segmentation through a dual monitoring system of mounted track cameras and autonomous drones. The system continuously scans tracks for damage, misalignment, or wear while also detecting and mapping railway assets like signals, signs, and switches. All track conditions and infrastructure objects are accurately mapped to their precise geospatial positions on a secure monitoring interface.
View
UMLify - Automated UML Diagram Generation from Code
Leverages static code analysis to automatically generate UML diagrams from uploaded source code. This tool streamlines software documentation, aiding developers in visualizing system architecture and improving team collaboration.
View
AgeLens AI - Image-Based Age Prediction
Employs advanced deep learning models to analyze facial images and accurately predict the age of individuals. The system leverages computer vision techniques to deliver reliable age estimations, with potential applications in identity verification, personalized marketing, and social analytics.
View
TourPlanner - Interactive Map-Based Itinerary Builder
Enables users to design personalized tours by selecting destinations, activities, and routes displayed on an interactive map. The system provides real-time suggestions and visualizes the itinerary, making trip planning intuitive and engaging.
View
Eviden - Video-Based Pipe Damage Detection
Integrates template matching to analyze video frames and identify regions of interest, which are then processed through a Convolutional Neural Network (CNN) to detect pipe damage. This system ensures efficient and accurate pipeline monitoring, enabling timely maintenance and reducing operational risks.
Closed Source
PaperLess - Distributed Searchable Document System
A scalable and distributed document management system built with independent Docker containers. It enables seamless indexing and searching of documents, ensuring high availability and flexibility for deployment across multiple environments. Perfect for organizations requiring efficient and distributed document retrieval.
View
Multi-Agent Camel Detection and Ranking System
A specialized program that detects camels based on the training set. Another model then checks if the camel is black. If all models pass, another model ranks the camel based on a dataset which has the price of camels at auction. Some features are e.g size, strength. The system automates the process of camel auctions, improving efficiency and fairness.
View
Eviden - Cable Length Detection
A program analyzes an image of a cable car cable. It calculates the length of the cable shown in the image. The program uses image processing and math for the calculation. It then displays the cable length as the result.
Closed Source
Style Transfer
A style-transfer application that can apply art styles or textures over new images.
View
GeoAI - Image-Based Country Detection
Employs advanced computer vision and machine learning models to analyze images and accurately identify their country of origin. The system leverages distinctive geographical, cultural, and architectural features, offering applications in geolocation, digital forensics, and tourism analytics.
View
RAG System - Retrieval-Augmented Generation
Uses word embeddings to identify context and then leverages large language models (LLMs) for accurate and minimal-hallucination responses. Ideal for precise, context-aware content generation.
Closed Source
VoiceGuard - Intelligent Trigger Word Detection
A Neural Network that detects the trigger word "activate". The system uses a combination of background audio, random word snippets, and "activate" snippets for training. When it hears "activate", it produces a chiming sound. The training data includes audio from various environments for robust detection.
View
JazzGenius - LSTM-Powered Music Generation
An LSTM network trained on a corpus of Jazz music to generate original compositions. The system predicts the next sound in a sequence, creating a chain of musical notes. Post-processing ensures the output adheres to musical conventions, producing coherent and pleasing Jazz melodies.
View