SAEEDA KANWAL

Passionate and Result-driven AI researcher with 8+ years of academic and applied experience in NLP, deep learning, and Bayesian modeling, seeking a PhD under the ERC-funded BALM project at Idiap. Skilled in transformers, variational bottlenecks, and knowledge graph embeddings, with hands-on work in Bayesian Transformers for Roman Urdu Sentiment Analysis. Strong programming (Python, PyTorch, TensorFlow) and mathematical background. Eager to contribute to interpretable,controllable attention-based architectures in collaboration with EPFL.

• Natural Language Processing
• Deep Learning and Representation Learning
• Bayesian Methods and Transformers
• Multimodal Information Fusion

 Roman Urdu Sentiment Analysis using Bayesian Transformers
GitHub: github.com/Saeeda-Kanwal/roman-urdu-sentiment-bayesian-transformers

• Developed a sentiment classification model for low-resource Roman Urdu using Bayesian Transformers.
• Integrated variational inference to capture prediction uncertainty and enhance model reliability.
• Preprocessed Roman Urdu text using custom tokenization and Byte-Pair encoding.
• Trained models with PyTorch and Hugging Face, outperforming baseline transformers.

Deep Learning Essentials with Keras,Tensorflow and Pytorch (Coursera)

• Published research on Intelligent Crowd Analysis in NED University Journal of Research (2018).
• Supervised GreenCity-AI, winning 1st place & PKR 500,000 at the Shell Tameer Awards.
• Finalist, INMIC conference 2024 Poster Competition
• Received PKR 90,000 ORIC funding for Hajj crowd flow analysis system.

• Speech-based Speaker Verification: Enhanced speaker verification using GMM-MFCC and CNN-RNN models, Achieved ~90% accuracy under real-world noisy conditions.
• StudyMateAI : AI-Powered Study Assistant, Created AI-based quizzes, note-taking, and video summarization tools., Leveraged Gemini embeddings, ChromaDB, Gemini 2.0 Flash technologies.
• Speech-to-Text Summarization in Roman Urdu: Built transcription and summarization system using Whisper and seq2seq models., Improved Urdu content accessibility for hearing-impaired and Roman Urdu readers.