Projects:

1. Visual Transformer Attention: Solved an implementation bug on a blog post and started implementing my own. Implemented Scaled dot product attention like the transformer paper on images like for example MNIST and Achieved Very good performance using very few parameters and less training time.

2. Sentence Interpolation using VAE and RNNLM: Implemented a Variational Auto Encoder based Language model for sentence interpolation using sentence embedding which can also be used for fill in the blanks. Used Quora question pair Analysis Dataset for this task.

3. Face Classification using CNN: Improved a CNN model with augmentations to get the best classification accuracy on Facial Images and won the competition.

4. Traffic Violation Detection: Implemented System to detect bike, helmet, human, License plate Recognize Numbers and characters from license Plate. Detect Bike and recognize License plate that have riders without helmet on

5. Object Detection and feature extraction using using MobileNet-SSD from Video (SSD.ipynb) : Used mobile net ssd for Object detection feature extraction used them for efficient tracking.

6. OCR Using CTC Loss for Bangla Printed Texts: Created in Reve Systems for Bangla data collection that can detect Bangla fonts better than open source Tesseract OCR from books for grammar and spell checker. Solved data scarcity problem with generated synthetic data with predefined font(as printed books) and complex text rendering library(Libraqm) with augmentation for robustness. Trained A Convolution Bidirectional (CRNN) model with Connectionist temporal classification(CTC) loss on the data.

7. Language Model for bangla from (GPT2, FASTtext, BERT(Multilingual), DistillBERT): Created language SOTA Language models using the data generated using the OCR.

9. Text Detection: Detected Text for recognition using EAST(Efficient and Accurate Scene Text Detection.

10. Facial Recognition:

• • • Website for REVE Systems Code (website link) : Created Face recognition system for Employee Security and Automatic attendance on Reve Systems.

    • ArcFace

    • FaceNet(Keras)

    • Openface(Keras)

    • Dlib

11. From Scratch Implementations:

• • • Implemented these algorithms from scratch and used them in various problem settings.

    • Decision Tree With Adaboost From Scratch

    • Recommender System (ALS) From Scratch

    • Neural Net with variable Hidden layers and activation functions from Scratch

    • PCA and Gaussian Mixture Models from Scratch

    • Template Matching (3 different Algorithms)from Scratch

    • Viterbi Algorithm for Channel Equalization from Scratch

    • Gaussian Naive Bayes from Scratch

    • Implemented my own grammar, lexar and parser for a simple calculator DSL language project in golang.