CV

Master's in Informatics

Technical University of Munich (TUM), Munich, Germany

• Thesis: Efficient Large-Scale Data Pruning Using Score Extrapolation (Supervisor: Prof. Stephan Günnemann), Grade: 1.0/1.0
• Major Coursework: Deep Learning, Statistical Learning, ML for Graphs and Sequential Data, Advanced NLP, Deep Generative Models, Computer Vision (Detection, Segmentation, Tracking), Computer Vision (Multiview Geometry) Quantum Computing, Tensor Networks, Deep Learning for Medicine
• Grade: 1.2/1.0 (Passed with High Distinction, Top 10%)

Bachelor’s Degree In Computer Engineering

Institute of Engineering, Pulchowk Campus, TU, Nepal

• Major Coursework: Theory of Computations, Data Structures and Algorithms, Operating Systems, Database Systems, Object Oriented Programming, Artificial Intelligence, Computer Organization and Architecture, Discrete Structure, Calculus, Data Mining, Linear Algebra, Probability, Statistics, Big Data
• Grade: 79.57% in aggregate (Top 5%)

Schmidt, S., Dhungel, P., et al. (2025). Effective Data Pruning through Score Extrapolation. In arXiv preprint arXiv:2506.09010. Link to Paper

Dhungel, P., et al. (2020). An Efficient Video Compression Network. International Conference on Advances in Computing, Communication and Networking (ICACCCN). IEEE. Link to Paper

Machine Learning Engineer

E.ON, Munich, Germany

• Architected end-to-end ML infrastructure on Azure using Terraform for infrastructure-as-code, enabling reproducible and scalable deployments with 99% uptime.
• Designed production-grade CI/CD pipelines (GitLab CI) with blue-green deployment strategies, achieving zero-downtime model updates and 30% cost reduction.
• Built comprehensive model lifecycle management using MLflow for experiment tracking, versioning, and performance monitoring.
• Led recruitment, onboarding, and mentorship of two junior data scientists.

Data Scientist (Working Student)

E.ON, Munich, Germany

• Led migration of on-premise data analysis to Azure, transforming a manual 3-day process into an automated daily pipeline, reducing time-to-insight by 66%.
• Architected Apache Airflow DAGs for automated data ingestion, preprocessing, model training, and evaluation with real-time anomaly detection.
• Developed a novel Robust-PLS algorithm for high-dimensional, correlated feature spaces, achieving a 7% improvement in MAE.
• Partnered with stakeholders to translate sensor physics into ML feature engineering strategies.

Data Scientist

Naamche Inc, Kathmandu, Nepal

• Developed end-to-end ML systems for real estate investment optimization, processing over $20M in property evaluations.
• Trained an ensemble of ARIMA, Prophet, and LSTM to forecast revenue with 14% hindsight error, improving portfolio returns by 12%.
• Architected a conversational real estate assistant using BERT-based intent classification and fine-tuned GPT-2 for natural language generation.
• Partnered with C-suite executives to define technical requirements and demonstrated ROI through bi-weekly demos.

Computer Vision Research Intern

NAAMII, Kathmandu, Nepal

• Designed 3D lung lesion segmentation pipelines for COVID-19 CT scans using U-Net and DeepLabV3, achieving results on par with SOTA benchmarks.

• Awarded the Deutschlandstipendium (sponsored by Amazon), granted to the top 900 out of 50,000 students, providing €300/month as financial support.

• Second Place: DandyHacks organized by the University of Rochester via Devpost

• Winner: A Day of Code

• Winner: Locus Codejam

• Represented Nepal in ACM-ICPC, Dhaka regionals.

• Winner: National Level Programming Contest

• Selected among top 5 students to represent Nepal at 47th International Physics Olympiad, Zürich Switzerland.

Efficient Large Scale Data Pruning with Score Extrapolation, Master's Thesis

• Proposed a novel framework for large-scale dataset pruning that eliminates costly full-dataset scoring. Our method extrapolates importance scores using geometric data structure (K-NN, GNNs), yielding competitive model performance with substantial end-to-end training time savings across multiple benchmarks and pruning techniques.

Video Compression Net, Undergraduate final year project

• A Deep Learning based video compression system that minimizes the number of bits required to encode video frames while also reducing distortion. The major objective of the project was to replace the conventional components in the pipeline of video compression with their neural network equivalents. The trained model equaled the compression ratio of the conventional system.

Covid-19 Emergency Chatbot

• A Viber bot, with more than 1100 subscribers, that caters to Covid-related queries in Nepali and English. With the addition of custom-made components like transliteration and NER, the chatbot was created using Rasa framework.

• Participated in the 3rd Nepal winter school of AI as teaching assistant in NLP

• Organized Locus Codejam

• Completed Deep Learning Specialization by Andrew Ng (Coursera)

• Participated in Hult prize 2019

• Organized and acted as a tutor on a competitive programming workshop