Close

Deval Shah

Machine Learning Engineer

About Me

Hello World. I am Deval Shah and I am in pursuit to understand and learn about the cognitive intelligence in machines. The idea of general intelligence in computers fascinates me. The visual perception domain of computers is where my interest lies. In recent years, the visual cognizance of machines has become more powerful due to the massive influx of research in the field of deep learning/computer vision. I believe that as computers become more intelligent and powerful and there will arise a great need to have an ethical code of conduct developed in those systems.


Currently, I am working as a Software Engineer/Researcher in Machine Learning at Uncanny Vision (AI based surveillance startup) in Bangalore, India.

Experience

Uncanny Vision

Deep Learning/Computer Vision


Automatic Traffic Congestion System (ATCS).

  • Developed a proprietary algorithm to improve the tracking of detected vehicles using base computer vision methods.
  • Worked on streamlining the end to end pipeline for the project.
  • Improved the stability issues of the system like memory leaks, crash due to excessive file writes, maximizing the CPU to GPU performance on the embedded device.

Optimization of Neural Networks.
  • Implemented a neural network pruning algorithm in PyTorch which led to 35%(reduction in training parameters) with 0.6% accuracy loss and 3x performance gains on Nvidia Tegra TK1 device

Optical Character Recognition(OCR).
  • Collected the Japanese character (kanji and hiragana) data (120k) with 165 classes by scraping the internet.
  • Developed a pre processing algorithm which increased the accuracy of training by a significant margin.
  • The custom architecture developed to train using the data and pre-processing algorithm outperformed the latest state of the art in Japanese OCR.

Dense Object Detection.
  • Worked on dense object detection problem to mitigate the false positive issue in two-stage detectors using Gaussian priors.
  • Developed methods to improve overall mechanism of training for binary classification in two stage detectors.

Automatic Number Plate Recognition(ANPR).
  • Worked on core algorithm optimizations for ANPR on low compute devices
  • Redesigned the post-processing algorithm for ANPR(Automatic Number Plate Recognition) system to efficiently queue unique vehicles in unconstrained environments with more than 97% accuracy.
  • Currently, working on vehicle re-identification across different ANPR edge devices for user verification.

Uncanny Vision

Research Intern

Developed azure based pedestrian detection tool for a client in initial phase of internship.

Automatic Number Plate Recognition(ANPR) team.

  • Wrote automation scripts in shell and python to enhance productivity of the data annotation team.
  • Built a sophisticated web scraper in python using selenium web driver to scrape more than 0.2 million unique Indian car images.
  • Created an automated test service on Azure for inference on stream data using object detection models and report mean average precision(mAP).
  • Programmed a fast linear motion tracking model in C++ to tackle repeat license plate issues in a single lane.
  • Optimized deep learning models using acceleration frameworks like OpenVino(Intel) and TensorRT(Nvidia).
  • Modified the architecture using LSTM's and improved the accuracy by 9%
  • Performed bug fixes for stability issues while running multiple instances of anpr system from same machine.

Whitepanda

Freelance Technical Writer

Curated articles,blogs, etc. on the basis of client's preferences on different techincal topics.

Towards Data Science

Data Science Writer

Wrote blogs on the topics related to Data Science domain. For eg : Activation Functions, Data Mining tools, etc. for TowardsDataScience publication.


Medium Profile

Education

Ahmedabad University

Fall 2014 - May 2018

Bachelor of Technology in Information and Communication Technology

Asia English School

June 2012 - March 2014

Higher Secondary Education

Projects

Automatic Number Plate Recognition(ANPR)

As an intern at Uncanny Vision, I worked on the development of ANPR system. It is primarily used for recognition of text on license plates. It has applications ranging from law enforcements, automated toll gates, gated community records etc.

**Can't share project link. It's confidential

Pedestrian Detection On Video

Pedestrian detection is an essential and significant task in any intelligent video surveillance system. It was developed using deep learning cnn based methods.

View Project

Face Recognition

Face Recognition system is developed using Eigen Faces algorithm. It is written in Matlab.It uses feature extraction and template matching based on Linear Algebra techniques.

View Project
Medium Article

Gesture Controlled Robot

The robot built is controlled by hand gestures. It was programmed on AtMega32 chip in C. It has various applications in military, warehouse management, evacuation etc.

View Project

Movie Tracker

A web application built for tracking movie information. It was written using Django REST API framework in Python.

View Project

Text Encryption Tool

Text encryption tool was built for encrypting text files using AES algorithm in Java along with GUI using JavaFX.

View Project

Phone Directory

Developed a phone based directory application in Java. The core features of CRUD(Create,Update,Delete) were built using efficient data structures.

View Project

Software


Frameworks


Technology


OS


Get in Touch