Imranul Ashrafi Email & Phone Number

Toronto, Ontario, Canada

- Improved core machine translation engine by using latest Transformer based architectures and language models.- Developed a paraphrasing based text generation model to generate 5X synthetic augmented data.- Developed a deep learning based bitext data cleaning methodology which uses language agnostic embeddings.- Implemented different Flask based APIs for generating back-translation and automated data extraction.

Zurich, Switzerland

- Architected and developed a customizable NLP module (similiar to AutoNLP) for long document texts which includes classifications, parsing, topic detection, summarization.- Developed a Parts of Speech based email parser to efficiently clean and extract text body.- Utilized contextual embeddings from HuggingFace Transformers and cosine similarity to extract keywords from text body. Used a lightweight pretrained language model (ALBERT) which reduced the model size by 72%.- Built a REST API using FastAPI for serving different models. Implemented prediction pipelines and optimized the neural networks by quantization for faster inference time.- Investigated and modified different libraries for improving app performance.- Active involvement in the code review process and maintaining best coding practices.- Writing unit tests for different reusable components on a regular basis.- Involved in developing frontend features of a chatbot software (aiaibot.com).- Led the frontend development of a recruitment management system.

Dhaka, Bangladesh

- Developed and deployed an OCR model and API on AWS that works on pictures taken by phone/web cameras, which produced scanned results as fast as 2 seconds per page along with preprocessing. Used CRNN architecture for the model and implemented it with Python and PyTorch.- Designed a custom text tokenizer and introduced different image augmentations to create a synthetic dataset of 1.2 million word images, which achieved an accuracy of 86% and reduced dependency on handcrafted real world dataset by 70%.- Participated in weekly paper reading sessions and presented 6+ state-of-the-art research works related to NLP.

Dhaka, Bangladesh

- Developed an efficient named entity recognition system by using a dataset of 72 Thousand sentences and achieved an accuracy of 91% on 9 different types of entities.- Created a dataset of 12 thousand sentences and built an ensemble model to generate questions automatically by predicting the relation between words, which achieved an accuracy of 73% on 10 different relation types.- Developed a text to image generation model using Generative Adversarial Networks (GAN). Used 5x5convolution instead of 3x3 to capture better image features.

- Developed a face anti spoofing model by extracting image noise with CNN and achieved an accuracy of 96%.- Generated synthetic data for a license plate detection model by using Generative Adversarial Networks (GAN) and used them to address dataset imbalance which improved the overall accuracy of the model.