Projects

Working with Dr.-Ing. Vincent Christlein on Enhancing Explainability in Handwritten Text Recognition (HTR) for Writer Identification using Transformer Attention and Register Tokens

Reproducing and benchmarking an advanced deep learning techniques on baseline HTR system with IAM dataset, achieving strong CER/WER scores and visualizing baseline attention maps for explainability analysis

Integrating register tokens into the Transformer encoder, improving attention clarity and interpretability while maintaining recognition accuracy, and demonstrated their potential for writer identification

Skills : Deep Learning, Explainable AI (Attention Maps, Register Tokens), Writer Identification, Model Benchmarking, Python, PyTorch, Hugging Face, Data Visualization

Source Code : To be published 

Working with Prof. Dr.-Ing. Jörg Franke and Christopher May on Design of multimodel synthetic data generation and annotation pipeline with vision model training

Designed an end-to-end synthetic data pipeline integrating Image generation models (Stable Diffusion, Flux), LLM-driven prompt augmentation (Ollama 2.0), and Grounded-SAM2 for auto-annotation, eliminating manual dataset creation

Trained YOLO models on fully synthetic datasets, achieving competitive detection accuracy across synthetic, cross-method, and real-world validation sets, proving pipeline scalability and adaptability

Pioneered a cost-efficient and automated approach to vision AI training, reducing dataset preparation effort by 80% and enabling rapid experimentation for diverse computer vision tasks

Skills : Generative AI, Vision Language Models, LLM Driven Prompt Engineering, Synthetic Data and Annotations Pipeline, Training and Evaluation

Source Code : To be published 

Worked with Prof. Dr.-Ing. Jan Dobschinski on Machine Learning Models for Offshore Wind Power Forecasting under Operational Constraints

Developed and benchmarked advanced forecasting models (XGBoost, MLP, ARIMA) using offshore wind farm data, improving short-term prediction accuracy under volatile wind conditions.

Engineered features from numerical weather prediction (NWP), historical production values, and curtailment signals, enhancing model robustness to grid-induced limitations.

Implemented data preprocessing techniques for curtailment handling, enabling models to distinguish between natural variability and grid congestion, thus supporting grid stability and market integration.

Skills: Time Series Forecasting, Machine Learning (XGBoost, MLP, ARIMA), Numerical Weather Prediction, Feature Engineering, Energy Systems Analytics, Data Preprocessing for Curtailment

Source Code : To be published 

Worked with Dr. Dario Zanca on Semi-Supervised Multimodal Tool Action Recognition Using Temporal Deep Learning Models

Developed BiLSTM and TCN-based deep learning models for tool action recognition, capturing temporal dependencies from multimodal sensor data (accelerometers, gyroscopes, magnetometers, microphones).

Implemented semi-supervised learning strategies (contrastive learning, mean teacher, pseudo-labeling), improving recognition F1 score from 0.33 to 0.70 with only 10% labeled data.

Conducted sensor fusion, ablation studies, and hyperparameter tuning, optimizing model performance for robust industrial monitoring and intelligent human-robot collaboration.

Skills: Semi-Supervised Learning, Temporal Deep Learning (BiLSTM, TCN), Multimodal Sensor Fusion, Data Analytics, Industrial Action Recognition, Python, PyTorch

Source Code/Demo : Here

Worked with Dr. Kaustubh Sakhare on Vehicle Detection in Aerial Images Using Traditional and Object Proposal Methods

Led a two-tier team to detect vehicles on the 2538 Vedai aerial image dataset using Object Proposal Methods (OPM) and selective search for precise object localization.

Conducted experimental analysis and feature extraction, boosting detection accuracy by 6% through implementation of an enhanced vehicle localization algorithm.

Skills: Object Proposal Methods, Selective Search, Feature Extraction, Vehicle Detection, Aerial Image Analysis, Python, Computer Vision

Source Code/Demo : Here

Worked on Gesture-Based Robotic Arm Control via Wireless Glove Interface - Human Machine Interaction

Developed a robotic arm system controlled by hand gestures through a wireless glove, enabling real-time motion recognition and execution of arm actions. 

Implemented Bluetooth-based communication (HC05) between glove and robotic arm, integrating Arduino UNO, DC motors, ADXL335 accelerometer, and gripper module for seamless control. 

Conducted experimental analysis and feature extraction, boosting detection accuracy by 6% through implementation of an enhanced vehicle localization algorithm.

Skills: Embedded Systems, Arduino Programming, Wireless Communication (HC05), Sensor Integration (ADXL335), Robotics, Gesture Recognition, Hardware Software Prototyping

Source Code/Demo : To be published