B M Abir

• CHOWA GIKEN Corporation

  Machine Learning Engineer

  2019 - Present

  • Creating custom Neural Network Architecture for supervised, semi-supervised Deep Learning Models using pytorch and tensorflow
  • Writing Metric Evaluation code to quantify model performance and accuracy
  • Curating Dataset by reviewing , modifying and restructuring so that it can be properly fed as input into custom neural network model
  • Write code for feature engineering by augmenting dataset
  • Hyperparameter tuning to improve accuracy
  • Deploy Machine learning models on Mobile Devices(Android)

  
  

• Monon AI Limited

  Lead Software Engineer

  2018 - 2019

  • Real time Video processing from IP Camera with OpenCV and python
  • Creating Multiple Face Tracker with OpenCV
  • Training Face Recognition models with Convolutional Neural Network(CNN) using Tensorflow
  • API endpoint consumption for analytics data
  • User Auth UI Design and Implementation
  • Docker deployment of python based image processing server in Google Cloud Platform(GCP)
  • Creating Number Plate Detection and recognition with Deep Neural Network(DNN)

  
  

• BRAC Institute of Educational Development, BRAC University [BIED,BU]

  Research Assistant

  2017 - 2018

  • Data Modeling and Database design
  • Augmenting Field level data collection into database systems
  • Real Time data processing and analytics

  
  

• Robotics Club of BRAC University

  President & Project Lead

  2015 - 2017

  • Co-ordinate between different Departments
  • Planning, organizing and executing event plans
  • Train members on the field of Robotics
  • Lead unique Research projects.

Education

BRAC University

Computer Science and Engineering

2017

Currently studying for undergraduate program in CSE. Took courses like introduction to database systems, Computer Architecture, Programing for the internet and Artificial Intelligence.

Wills Little Flower Higher Secondary School

O Levels & A Levels

2010 2012

Completed my O-levels with 6 A's and A-levels with 2 A's and 1 B's grade. My Core Subjects were Maths and Science.

Projects

Tran Dao

A COVID-19 Relief Distribution app to record and suggest where more relief is needed

2020

FaceLens

Real time Face Recognition Platform

2019

I have experience with building Real time Face Detection and Recognition Platform. I was responsible for Leading a group of developers where i designed and implemented a Video Processing pipeline to detect face using a Deep Neural Network Based Detector and tracker. Then each Detected Face was recognized with a Convolutional Neural Network Based Classifier which i have trained using CelebA Dataset. The Classified face was recognized from a pool of Registered Face from database. I was also responsible for designing and overseeing a restful Api server using python to deploy this detection, recognition and querying of faces from database. The Platform is deployed as a fully scale-able cloud service with detection of faces computed at Edge( For Decentralized Camera Network).

BIED Survey

Data collection, analytics and visualization

2018

A webApp to collect survey data. The respondents fills up a form and submit their answers. An admin dashboard is present to add, edit and delete the questions of the survey. The results of the survey data is displayed with graphical pointers. And the results can be exported in varieties of formats. The collected data is then used to train and predict pedagogical analytics model with regression.

Bangla Handwritten Character Recognition

Using Convolutional Neural Network

2016-2017

Handwritten character recognition from a natural image has a large set of difficulties.Bangla handwritten characters are composed of very complex shapes and strokes.Recent development of deep learn- ing approach has strong capabilities to extract high lavel feature from a kernel of an image.This Project will demonstrate a novel approach that integrates a multilayer convolutional neural network followed by an in- ception module and fully connected neural network. The proposed ar- chitecture is used to build a system that can recognize Bangla charac- ter from different writers with varied handwriting styles. Unlike previ- ous handcrafted feature extraction methods, this CNN based approach learned more generalized and accurate features from a large-scale train- ing dataset. 1,66,105 training images of Bangla handwritten character of different shapes and strokes has been used to train and evaluate the per- formance of the model. Thus allows a higher recall rate for the character in an image and outperforms some current methodologies.

Mongol Tori

A Mars Rover Project by BRAC University

2015-2016

I was the lead Software Developer for the Mars rover Robot. It had the capability of working remotely by communicating with the control station and used GPS to track the location of its terrain traversing.

RFID DoorLock with WebApp

An access Control System for ROBU

2014-2015

• Members Management
• RFID card Database

pre-University Consultation System

A Self-Assessment Software for college grad. Students

2015

• Student Self-Help
• Finds what to study in university
• Communicates with professors

Smart DoorLock

App to control your door from any location

2015

• Android Control
• DoorLock

Publications

Bangla Handwritten Character Recognition With Multilayer Convolutional Neural Network

Advances in Data and Information Sciences. Lecture Notes in Networks and Systems, vol 39. Springer, Singapore

Nov 17, 2017

Abstract. Handwritten character recognition from a natural image has a large set of difficulties.Bangla handwritten characters are composed of very complex shapes and strokes.Recent development of deep learn- ing approach has strong capabilities to extract high lavel feature from a kernel of an image.This paper will demonstrate a novel approach that integrates a multilayer convolutional neural network followed by an in- ception module and fully connected neural network. The proposed ar- chitecture is used to build a system that can recognize Bangla charac- ter from different writers with varied handwriting styles. Unlike previ- ous handcrafted feature extraction methods, this CNN based approach learned more generalized and accurate features from a large-scale train- ing dataset. 1,66,105 training images of Bangla handwritten character of different shapes and strokes has been used to train and evaluate the per- formance of the model. Thus allows a higher recall rate for the character in an image and outperforms some current methodologies

Media Exposers

BBC Bangla- Solving Traffic Problem with AI Solutions

Jetro- AI startup from Hokkaido University establishes development base in Bangladesh