I’m a Machine Learning Engineer and AI Researcher based in Germany 🇩🇪, with a strong focus on deep learning for sensor data, biosignals, and computer vision systems. I bring expertise in developing robust AI pipelines—from raw data acquisition to deployment—for both research and industry applications.

⚡ I’m passionate about applying AI to real-world domains like marine robotics, healthcare, and multilingual NLP.

⚡ I regularly explore topics like domain generalization, generative modeling, and signal processing.

⚡ I’m currently working on AI validation systems for marine sonar data and fine-tuning LLMs for domain-specific translation tasks.

• Languages: Python, C++, Java
• Deep Learning: PyTorch, TensorFlow, Keras, Scikit-learn
• Computer Vision: OpenCV, CNNs, GANs, Object Detection
• Robotics: ROS, Sensor Fusion, Motion Planning
• Cloud & MLOps: AWS, Azure ML, SQL, REST APIs, Docker, Kubernetes
• Others: GitHub Actions, Linux CLI, Bayesian Learning, NLP

• Built a domain-specific translation platform by fine-tuning LLMs using LoRA (Low-Rank Adaptation).
• Designed full data pipelines and deployed the system as a RESTful service using Azure Functions.
• Tech: PyTorch, Hugging Face, AWS, SQL, Azure, Python

• Developed end-to-end AI pipelines for sonar data using U-Net and GANs for data augmentation.
• Integrated LiDAR + camera data for a real-time lane-following robot with obstacle avoidance.
• Tech: PyTorch, OpenCV, GANs, Docker, C++

Built a real-time sonar imaging system for object detection in marine environments.

• Leveraged Generative Adversarial Networks (GANs) to synthesise and augment sonar imaging data, enhancing training diversity and robustness.
• Engineered an end-to-end AI pipeline: from raw sensor acquisition to real-time inference for sonar-based object detection.
• Designed modular components for processing sonar signals and integrated them with LiDAR and camera sensors for a multi-modal robotic system.

Impact: Improved detection performance in safety-critical marine environments with limited training data.

Developed an autonomous lane-following robot using a TurtleBot platform, combining sensor fusion and classical image processing.

• Integrated LiDAR and camera data on TurtleBot for accurate perception of the environment.

• Implemented lane detection using image processing techniques for visual guidance.

• Utilized LiDAR-based localization to support real-time path planning and obstacle avoidance.

• Combined ROS-based modules for sensor control, motion planning, and control loops.

This project evaluates generalization of emotion classification in biosignals across domains using three techniques:

1. Empirical Risk Minimization
   Based on Vapnik's theory (1998), minimizing sample-based error while considering distribution ( D(S) ) over domain ( X ).

2. Multitask Bayesian Network
   Modeled complex relationships in biosignal data with uncertainty modeling (inspired by Dissanayake).

3. Convolutional Neural Network
   Feature learning from time-series transformed signals, based on architectures such as those from Ballas et al.

Applying machine learning to industrial settings often faces the challenge of limited and irregular sensor data. This project tackles automatic feature extraction for time-series data from real-time accelerometer signals monitoring the
comfort-closing of car doors.

• Explores three approaches:
  1. Pre-trained models on signal data
  2. Frequency-time domain analysis
  3. Image-transformed signal features using CNNs

• M.Sc. Autonomous Systems – Hochschule Bonn-Rhein-Sieg
• B.Tech Automation & Robotics – B.V. Bhoomaraddi College

• LinkedIn
• GitHub
• Reach me at: janhavi.puranik1995@gmail.com

🧩 Always open to freelance or collaborative projects in ML research, robotics, and intelligent systems!