MSc Artificial Intelligence and Robotics · University of Technology Nuremberg
I build practical machine-learning systems across computer vision, geospatial intelligence, generative AI, data engineering and robotics. My work focuses on reproducible experiments, real-world robustness and deployable applications—not only model training.
- Robust computer vision under compression and distribution shift
- Efficient deep learning for deployment-constrained systems
- Geospatial machine learning and urban-change analysis
- LLM applications, RAG and agentic workflows
- Reproducible research pipelines and evaluation
| Project | What it demonstrates | Stack |
|---|---|---|
| Deepfake Detection Under Real-World Degradation | Leakage-safe benchmarking across compression, resizing, blur and compound social-media transformations | PyTorch, EfficientNet-B0, Grad-CAM, pytest |
| European Image Geolocation | Sub-5M-parameter hierarchical geolocation with spherical decoding and offline inference | PyTorch, RegNet-Y, GeM, scikit-learn |
| Nuremberg Land-Cover Intelligence | Sentinel-2 land-cover composition and change estimation over a 100 m urban grid | LightGBM, GeoPandas, Rasterio, Streamlit |
| Redshift Streaming Analytics | Reproducible streaming analytics with local and cloud storage paths | Kafka, DuckDB, Redshift, Docker, Streamlit |
| Dual-Adapter LoRA Style Tuning | Parameter-efficient Stable Diffusion adaptation with custom-token gradient masking | Diffusers, PEFT, PyTorch |
| SpeechText | Browser audio transcription and persistence on Google Cloud | Node.js, Speech-to-Text, Firestore, App Engine |
- Deepfake robustness: evaluated lightweight detection across realistic degradation tiers with ROC-AUC, calibration error and bootstrap confidence intervals.
- Image geolocation: 150.67 km validation median error using a single 4.63M-parameter model trained from scratch.
- Land-cover estimation: built-up proportion spatial CV of 0.0489 MAE, 0.0979 RMSE and 0.9146 R².
- Research publication: Deep Gesture Interpretation for American Sign Language, using EfficientNet-B0 for 29-class recognition.
Languages: Python, C++, C, JavaScript, SQL
Machine learning: PyTorch, TensorFlow, Keras, scikit-learn, LightGBM, NumPy, Pandas
Computer vision: EfficientNet, ConvNeXt, RegNet, OpenCV, Grad-CAM, robustness evaluation
LLM systems: Hugging Face, LangChain, RAG, FAISS, agentic workflows
Data engineering: Kafka, DuckDB, ClickHouse, PostgreSQL, dbt
Applications and tooling: Docker, Streamlit, FastAPI, React, Git, Linux
Previously worked as an AI and Automation R&D Intern at ICAR–NBSS&LUP, developing Python APIs and data-access automation. I am currently pursuing an MSc in Artificial Intelligence and Robotics at UTN and am interested in working-student and research opportunities involving applied AI, computer vision, speech technology and robotics.
- Measurements that reflect deployment conditions
- Clear baselines and honest limitations
- Leakage-safe evaluation
- Reproducible commands and configurations
- Small, understandable systems before unnecessary complexity