Experience

Machine Learning Engineer, Sensus Labs

Mumbai, Delhi
Aug 2018 - Sep 2019
In Sensus labs an IIT Bombay alumni, I’d worked as a Full-stack Machine Learning Engineer for more than a year As an ML engineer: Here, I've worked on creating complex Machine Learning pipelines in order to solve among the toughest Computer Vision & other Machine Learning based problems like Fraud Detection, Demography Analysis, Cloud Kitchen monitoring, Footfall analysis, etc by using computer vision & deep learning to achieve an Industrial Automation and deployed at scale using TF-serving,Kubernetes & Flask

As a Machine Learning Instructor: Worked as an Instructor & Lead for Data science Interns, assigned & helped them to tackle day-to-day tasks related to Machine Learning

Machine Learning Engineer, Omdena

Remote ( USA )
Aug 2018 - Current
Omena is a global collaborative platform where AI enthusiasts from diverse backgrounds build AI models and solve real-world social problems
I got accepted among the 10k+ applications in a team of 32 engineers globally to build an AI solution regarding the Prevention of Sexual Harassment in India on Privately collected Satellite Image & Statistical Data by Omdena itself by building a geographical model to detect safety of regions

Computer Vision Research Intern, Sensus Labs

Mumbai
Jun 2018 - Aug-2018
Worked on creating a custom Object Detection model using Keras & Tensorflow to monitor the footfall on various stores at a given time.
Work Highlights:
Image Pre-processing: Worked in Pre-processing image making it suitable for the input for the model & creating bounding boxes & masks
Experimenting with different models: I was given a task during my internship is to use different models and test Top-1, Top-3 & Top-5 accuracies on pre-trained CNN models on Image net & COCO dataset

Professional Full Stack Web Developer

Mumbai
Jan 2017 - Jun 2018
Worked as a Full stack web developer and created and contributed to various web applications using technologies like
nodeJS ,MongoDb , SQL , React ,bootsrap ,Express , RestAPI ,etc

Skills

ML & Data science skills

Computer Vision
NLP
Statistical Modelling
Statistical Inference

Languages

Python
Javascript

Machine Learning Libraries

Scikit-Learn
XGBoost
LightGBM
OpenCV
NLTK

Deep Learning Libraries

Keras
Tensorflow
Pytorch
Theano

Databases & Data Pre-processing:

Pandas
Numpy
SQL
Mongo DB

Software Stack:

Docker
AWS
Google Cloud
Heroku
Linux
Flask
NodeJS
REST services
Git

MOOCS

Stanford Online: Convolutional Neural Networks for Visual Recognition(CS231n)
Stanford Online: Deep Learning (CS230)
Coursera: Deep Learning specialization
Coursera: Intro to Deep Learning by national research USE
Coursera: Intro to Machine Learning By Andrew NG
Udemy: Data Science Bootcamp with Python
Udemy: Complete Full Stack web developer bootcamp
Coursera: Advanced Machine Learning and Signal Processing(Enrolled)
Udemy: Big Data with Pyspark(Enrolled)

Volunteers

Machine Learning Developer (Omdena)
I've participated in one of the Omdena challenge as an machine learning Engineer to build AI solution to help the prevention of sexual harassments cases in India
Lead Software Developer (Palghar District police station)
I voluntarily participated In this project which helps in better management, I wrote the whole backend of a NodeJS based full-stack web application which manages profiles of different level of police officers and their transfer status with a high level of authentication security
Lead Software Developer (Computer Society of India)
I worked on this non profit projects as a lead full stack Web developer and created a web application that helps with all the events created by CSI and registrations of all candidates.

Projects

Designing efficient Video Fraud analysis Pipeline in DL & CV

Complete overview of the pipeline that I created to run Video Fraud detection model more efficiently
- Total Video classification + Object detection inference time = 13 sec ~ 15 sec
- Flaw : Inference time was really long
- Reason : I was loading separate CNN model into memory to extract CNN features of video frames in CNN+LSTM model
Pipeline
- Total Video classification + Object detection inference time = 2 sec ~ 4 sec
- Changes : IInstead of Loading the CNN model in memory at only the video classification section we are extracting image features at the initial step.
Pipeline

Github :

Video Classification using CNN + LSTM

This is one of the project that I implemented from Multiple Research papers to classify Video So my intuition here is basically converting video frames in constant Sequence Inputs and train it on many to one RNN model to classify video
Steps:
- Breaking Video into frames and Pre-process each frames
- Extract a CNN features by passing image data into pre-trained CNN model and saving the output of (N-1)th layer of network
- Converting Extraced Features into single sequence input
- Train a LSTM classifier network based on this sequence
Pipeline

Loss Plot

Accuracy achieved : 80%

Github :

Custom-Mask_RCNN-for-detecting-states-of-object

This project is part of our trial and error Pilot to detect Fraudulent Activity which we were working for one of our clients,here model will get input will give states of objects

Here I trained a Mask_RCNN model on our custom data to detect the occupancies of different object(eg. Occupied chair , Unoccupied chair)

Results Github :

Reddit RoastMe_BOT using Deep Learning

In this project, I have created the Reddit Roastme data-based model that generates text given the input image & text using Deep Learning.

I have trained a 2 LSTM model with sequential text data (multiple comments) & title with its corresponding CNN extracted feature data of that image post

To Do this , I combined Chatbot & Image Captioning Model into single Model

I cannot explain whole implementation just here, so I have written a medium post Link in lowest technical detail possibleb

Convergence Graph Medium post of this link

Github :

Supervised-machine-learning-AI--bot_in_GTA_SA-game

In this project i have created ai bot of player & auto driving car using convolution neural networks and open CV

In first part just using OpenCV I applied various cv filters to detect lanes

In second part(using deep learning) I trained my own gameplay and keylogs in CNN classifier

Working Demo(Self driving car in FPS mode)

Github :

YOLO-Tensorflow_serving_Flask

In this project ,I created end-to-en object detection pipeline which takes image or video as a input and returns class names & object bounding box location and deployed using TF-serving & Flask

Process I converted weights file of keras darkflow tiny_yolo model into frozen pb format for tf serving and hosted in tf-serving docker and
created external flask api backend which is using tf-serving REST api ,

this backend takes base64 format of image and outputs a JSON output

Frontend CLI version
Deployment map

Github :

Omdena_Harrasment_prediction_competion

This project is part of my omdena global competition, here we are given an anonymous crime data collected by safecity consist of Lat & long and incident record, I am creating machine learning to predict the safety of region.
The features that we are dealing with are really granular in nature but similar(present of schools, street roads, building rooftops, etc)
So I think if we could train a model to identify the similarity between images in terms of recurrent features then it will be much easier to identify the safeties of the given input map image.

Embedding

Feature Engineering:
note that this feature alone won't able to create feature ,so I am planning to use other numerical & categorical data

Features(as of now):
- UnProcessed Satellite image
- Bounding Box locations of buildings for spatial feature extraced my pre-trained MASK_RCNN.
- Detecting presence of (Schools,Police Stations , Companies , Hospitals) using OCR on non-sattelite image.
- Statistical data like location wiseCrime conviction rate,Literacy rate ,etc. from data.gov.in
Mask RCNN Part :

To identifying the buiding positions building , I am planning to use pre-trained mask-rcnn model used by Neptuneml on sattelite images

OCR part
Identifying the presence of Schools,Police Stations , Companies , Hospitals,etc by eunning on this version of image
Github :

Palghar-Poilice-Transfer-System-Managment

In this project I wrote the whole backend on my own of a NodeJS based full-stack web application which manage profiles of different level of police officers and their transfer status with high level of authentication security

Github :

Detecting Fraud using Deep Learning

I've been part of this project during my time at sensus labs, here we detect fraudulent activities happening at the reception across all the OYO based hotels

We trained spatial-temporal data of videos on CNN-LSTM based for activity classification.

Github :

Mini Projects

Food Quality Monitoring
In this project, we created a custom image classifier & object detection model by training on different components of food dishes combination of hoi foods on mobilenet model and deployed in Raspberry Pi for HOI Food Kitchen to maintain their quality assurance

Demography Analysis
This project is our part of creating Demography analysis at one of the biggest food market chain HMR in Russia, we used Face Detection module & Dlib to detect Age & Gender of customers standing at the reception.

Generating Abstract Art using Deep convolutional generative adversarial networks[DC-GANs] :
In this project, I have generated abstract by training DC_GAN model with approx 8500 painting
Github :

Predictive Model using Ensemble Learning on XGBoost, Keras & Scikit-Learn : :
This project is part of my participation in one of the kaggle competition , here I created Regression Model by combining many different regression Models like RFRegressor,Keras MLP,XGBoost, etc and combining those using Stacking ensemble learning techniques on numerical & categorical data by doing extensive feature engineering by extracting Statistical Features.

ResNet implementation in TensorFlow Keras :
In this project ,I implemented residual network from ground-up to get a better intuition about how Resnet & CNN works.

Generating text using LSTM :
In this project I trained a generative LSTM network shakespeare writings to generate text from it.

Implementing 3D convolutional network from Research paper :
In this project I crerated 3d Conv model to classify video action from a 3d conv based research paper.