KI ML LLM Entwickler Python GPT RAG Data Science Computer Vision OpenAI Deep Learning

Project: Smart Trunk Opener

As a Senior Data Scientist, I led the development of a Smart Trunk Opener project (joint project of infineon and intive), which aimed to accurately predict kick gestures for hands-free trunk opening using the BGT24***** radar based on the Doppler effect and using Neural Networks on embedded devices.
- Developed algorithms to detect abnormal signals and validate them as kicks using Python and Matlab
- Created NN models for better predictions, and optimised them using TensorRT for efficient use on embedded devices.
- Improved data reception and pre-processing on the radar using the C programming language.
- Compiled complete production code as C code using Atmel Studio and deployed it on the embedded system.

Tech Stack: Python, C++, C, Neural Networks, NumPy, Pandas, Pytest, Pytorch, Linux, Bash, Git, Matlab, Ifxdaq (Infineon Data Acquisition), Matplotlib, Time Series analysis, ARM Cortex, ARM Socket, RTOS, ATmega, Atmel Studio.

Project: Bacteria Detection

Development of a bacteria detection project using microscopic images, achieving 99% accuracy in detecting the location and number of bacteria.
- Trained yolo5 deep neural networks to extract relevant bacteria bounding boxes using IoUs.
- Accelerated the AI on Jetson Nano by converting to TensorRT (CUDA) and developed the final C++ code interference with pre- and post-processors.
- Optimized the yolo network decoder from Python to C++ and programmed tensors directly on the GPU using CUDA C++.
- Reduced the AI utilizing Knowledge Distillation techniques to achieve high performance processing on the Jetson Nano.

Tech Stack: Python, C++, C, Neural Networks, Machine Learning, Object Detection, yolo, OpenCV, Pytest, Pytorch, CUDA, Numpy, Matplotlib, TensorRT, Git, Jupyter Notebook, OOP, Linux/Bash, AWS, Sagemaker, PyTest, Unittest, Jetson Nano.

Project: Robot Manipulator Optimization

Determination of the optimal trajectory for the robot arm. I have developed a custom neural network to solve this optimization problem using physically modeled helper computations.
- Conducted research on optimal control problems via neural networks.
- Generated simulated data based on classical mechanics, performed data processing and normalization, and converted
   original Matlab files to Numpy.
- Constructed LSTM networks
- Enriched data with cubic splines and implemented objectoriented programming to produce productive and efficient
  code.

Tech Stack: Neural Networks (Recurrent Nets), Supervised Learning, Python, C++, Pytorch, Cuda, Numpy, Matplotlib, Torchdiffeq, MLFlow, Tensorboard, Autograd, Integrate, Torchsummary, Scipy, Torchcubicspline, Git, Jupyter Notebook, OOP, Pytest.

Project: Breast Cancer Detection on MRI

Development of a neural network-based algorithm to detect, classify and segment breast tumors in mammography X-ray images to improve radiologists' performance in breast cancer screening.
- Used segmentation techniques to train Mask R-CNNs to detect the tumors.
- Improving the inferenc via pretraining of the Mask R-CNN with the bounding box mask of the Yolo detector. Then
  training on qualitatively better but few masked data.
- Integrated sensitivity and precision metrics, and achieved stateof- the-art results.
- Deployed the finished code on AWS endpoint for professional level breast cancer detection.

Tech Stack: Python, C++, Mask RCNN, Segmentation Networks, PyTorch, NumPy, Machine Learning, OpenCV, AWS, Pandas, Git, Jupyter Notebook, Unittest.

Project: Fall Detection

Development of Fall detection Machine Learning Algorithms for a Smart Vision Assistant.
- Collect data of human skeleton in various fall scenarios via motion capture and skeleton detection algorithms.
- Balance the dataset with fall and non-fall scenarios to ensure that the machine learning algorithm can distinguish between normal and abnormal movements.
- Feature Extraction: relevant features extracted from skeleton data to represent the movement patterns: joint angles, joint velocities, and acceleration
- Data saved on AWS S3 bucket.
- Extracted features used to create matrices which were fed into a machine learning algorithm.
- Used AWS AutoML to find best ML model.

Tech Stack: Python, Scikit-learn, OpenCV, Pytorch, Image-Pose, AWS, Sagemaker, AutoML, AutoPilot, Cuda, Linux, Git, Pytest.

Project: Voice ReID, Voice to Text

Development of speech recognition system that can identify the speaker and transcribe the language from audio data.
- Audio data divided into 4-second segments sampled at 16kHz.
- Feature extraction techniques on audio data, such as MFCCs, Mel-scale spectrogram, chromagram, spectral contrast, and tonnetz, based on STFT, utilizing Kaldi.
- Utilizing Scikit-learn for feature clustering.
- Utilizing FFT (Fast Fourier Transformation) for denoising and feature extraction.
- Creation of VoiceReID custom model via PyTorch and TensorFlow.
- NLP algorithms for speech-to-text transcription, utilizing part-of-speech tagging and word sense disambiguation.
- Utilizing Gensim and NLTK for text summarization, tokenization, stemming, lemmatization, part-of-speech tagging, parsing
- Applied named entity recognition to identify named entities in text data and categorize them into predefined classes

Tech Stack: Python, pyannote, Kaldi, NLP, AWS Sagemaker, EC2, S3, Lambda, NumPy, TensorFlow, LSTM, Scikit-learn, NLTK, Spacy, Gensim, CoreNLP, Transformers, Hugging Face, Pandas, SQL, FastAPI, Git, Bash, VS Code, Unittest, Poetry.

Project: Box Optimization
Minimizing the amount of empty space in packages. This reduced not only CO2 emissions but also transport costs. My solution achieved a 10% improvement in packing efficiency.
- Utilized data sampling, dimensionality reduction and data approximation techniques to reduce the data size.
- Modelled the packages as matrices to rotate and modify easily via numpy.
- Developed evolutionary algorithms with efficient mutation and breeding methods as tensor operations on CUDA to
   have very fast parallel breedings.
- Benchmarked the results using MILP optimiser.

Tech Stack: Python, Numpy, Torch, Cuda C++, Pandas, SQL, Pytorch, Cuda, Cudnn, AWS, Sagemaker, Docker, Git, Jupyter Notebook, OOP, Clustering, NP-hard problems.

Project: Invoice Recognition, OCR
Development of an OCR for invoice recognition software that checks for invoice errors.
- Built an efficient SQL command search function to simplify the work.
- Used OpenCV to apply image recognition techniques.
- Implemented deep learning techniques for text recognition on invoices.
- Developed deep auto-encoders for data dimension reduction and noise reduction.
- Used Convolutional Neural Networks to improve logo recognition.

Tech Stack: Machine Learning, NLP, Python, Tensorflow, Numpy BigData, PyTesseract, Git, Bash, AWS, VPN, SQL.