Some things I know in life! 😎

Hi Guys. I am a passionate engineer interested in Deep Learning, Computer Vision, Photogrammetry, and Generative AI. I love to develop deep learning based computer vision algorithms that can have a real imapct in terms of performance, generalizability, and outreach of its solution. Developing the algorithms for surviellance industry for more than six years have given me the “deep” understanding behind DNNs(Pun Intended !) and I will cherish and contine my chase for Applied AI in solving real world problems. My Masters thesis was based on Highway Traffic Analytics where I developed a real-time high performance analytics pipeline optimized for highway networks.

I also like to write the technnical explainations of the Deep Learning Research Papers. Dont forget to visit on my Medium Page.

Apart from the academica, I like to play cricket 🏏, football ⚽, and badmintion 🏸. I also like to read books related to sci-fi, thriller, cosmology, and astrophysics. Not to forget, I am a huge Real Madrid Fan. HALA MADRID !.

Last but not least, I am a Pokémon freak. I judge a person based on which starter pokemon they choose. I choose Charmender!

---

Table of Content

• Education 📚
• Courses and Electives ✍
• Languages, Frameworks, and Tools 🔏
• Experience 🧠
• Patents 📜
• Open-source Contributions 💻
• Blog Posts ✍️
• Projects 📝

---

Blog Posts

July 16, 2020

DetectoRS — A Comprehensive Review

State-of-the-art object detector from Google Research using Recursive Feature Pyramids and Switchable Atrous Convolutions achieving 54.7% mAP on COCO.

Object Detection Google Research

June 28, 2020

MixConv — Mixed Depthwise Convolutional Kernels

Understanding Mixed Depthwise Separable Convolutions from Google Brain for efficient ConvNets on edge devices.

CNN Architecture Google Brain

June 18, 2020

RegNet Architecture from Facebook AI

Simple, Powerful, and Fast — Understanding the RegNet architecture designed for efficient neural network scaling.

Neural Architecture Facebook AI

June 12, 2020

Understanding Attention Modules: CBAM and BAM

A quick read on Channel and Spatial Attention mechanisms that help CNNs focus on important features.

Attention CNN

May 28, 2020

YOLOv4 — Final Verdict

Complete breakdown of YOLOv4 architecture, training strategies, and why it became the state-of-the-art detector.

YOLO Object Detection

May 23, 2020

YOLOv4 — Bag of Specials

Deep dive into YOLOv4 Bag of Specials — attention modules, feature aggregation, and activation functions.

YOLO Deep Learning

May 2020

Understanding Focal Loss

How Focal Loss addresses class imbalance by down-weighting easy examples and focusing on hard negatives.

Loss Functions Object Detection

April 2020

Object Detection — A Complete Guide

Comprehensive introduction to object detection in Machine Vision — from fundamentals to modern architectures.

Object Detection Tutorial

---

Education

College/School	Degree	GPA/Percentage
Lassonde School of Engineering, York University	Masters in Computer Science - Thesis	3.9
Ahmedabad University	BTech. Information and Communication Technology	3.61
Swastik Sindoor	Higher Secondary Board	92%
Swastik Sindoor	Secondary Board	85%

Courses and Electives

• Introduction to Blockchain
• High Performance Computing
• System Design for Societal Problems
• Computer Vision
• Advanced Data Structures and Algorithms
• Machine Learning
• Cloud Computing

• Data Analytics and Visualization

  Core Courses

  • Basic Electric Circuits
    
  • Calculus and Differential Equations
    
  • Communication Skills
    
  • Introduction to Computer Programming
    
  • User Centered Design
    
  • Digital Design
    
  • Discrete Mathematics
    
  • Electromagnetic Theory
    
  • Enviornmental Studies
    
  • Object Oriented Programming
    
  • Computer Organization
    
  • Data Structures and Algorithms
    
  • Ethics
    
  • Linear Algebra
    
  • Signals and Systems
    
  • Analog and Digital Communication
    
  • Database Management Systems
    
  • Embedded System Design
    
  • Probability and Random Processes
    
  • Computer Networks
    
  • Operating Systems
    
  • Wireless Communication
    
  • Digital Signal Processing
    
  • Integrated Circuit Device and Technology
    
  • Software Engineering
    

---

Languages, Frameworks, and Tools

Languages I am most confident in

• C
• C++
• Python
  Other Languages

  • Java
    
  • SQL
    
  • HTML
    
  • CSS
    
  • Ruby
    
  • LateX
    
  • Markdown
    
  • JavaScript
    

Frameworks and Tools I can definitely work on

• PyTorch
• TensorFlow-2.0
• TensorRT
• OpenVino
• MLFlow
• Docker
  Other Tools/Frameworks

  • MatLab
    
  • GCP
    
  • MySQL
    
  • Darknet
    
  • OpenCL
    
  • GIT
    
  • Docker
    

TODO List

• Nvidia’s Deepstream and GStreamer
• Grafana
• Prometheus
• Elastic Search
• Kubernetes
• GoLang
• KubeFlow

---

Experience

• Machine Learning Researcher at Cloudastructure Inc. (Aug 2025 - Present) - Remote, India
  
  • Integrated attribute detection directly into the YOLO architecture, eliminating downstream classifiers and achieving a 4x speed boost.
  • Built a camera mapping framework leveraging SfM, photogrammetry, MCMO tracking, and ground-plane alignment—cutting false alerts by over 50%.
  • Developing an Auto-Annotation pipeline using Prefect, Haystack, and VLMs to enable continuous labeling, online training, and data generation for downstream systems.
• Computer Vision Engineer at Trans-plan (May 2023 - Aug 2025) - Toronto, Canada
  
  • Worked on advanced traffic analytics using drones and infrastructure cameras for real-time highway monitoring.
  • Improved Python workflows to process and georeference dynamic drone footage using GDAL, rasterio, and shapely.
  • Delivered actionable data products including traffic density maps, vehicle flow rates, heatmaps, and inbound/outbound counts to support better traffic management.
• Machine Learning Engineer at Ontario Ministry of Transportation (MTO) (June 2023 - July 2025) - Toronto, Canada
  
  • Built real-time vehicle counting with lane-level classification and adaptive clustering.
  • Ported to NVIDIA DeepStream with custom plugins for counting, tracking, and messaging.
  • Deployed plug-and-play edge unit with IP camera and simple management interface.
  • Designed secure 4G data transmission (0.2ms latency) and live traffic dashboard.
  • Containerized stack with recovery, health checks, and data caching for zero data loss.
• Machine Learning Research Assistant at Elderlab, York University (Sept 2022 - July 2025) - Toronto, Canada
  
  • Built & deployed a real-time day/night highway analytics stack (detect–track–count) on Highway 401 with Ontario MTO, running on DeepStream + TensorRT edge devices for 24/7 lane, speed, density, and anomaly metrics.
  • Optimised vehicle tracking to <1% daytime and 1.4% nighttime counting error in rain, snow, and fog.
  • Designed unsupervised lane-assignment algorithm via backprojection on ground plane by simulating the oriented bounding boxes around vehicles in real-time.
  • Created an unsupervised domain-adaptation object detection training pipeline using astronomical, nautical, and civil dusk/dawn pseudo-domains, boosting nighttime detection mAP.

• Graduate Teaching Assistant at Lassonde School of Engineering, York University (Sept 2022 - July 2025)
  

  Semester	Course
  Jan 2025 – Apr 2025	EECS1730 - Building Interactive Systems
  Jan 2024 – May 2024	EECS2101 - Fundamentals of Data Structures
  Sep 2023 – Dec 2023	EECS1015 - Introduction to Computer Programming
  Jan 2023 – Apr 2023	EECS1720 - Building Interactive Systems
  Sep 2022 – Dec 2022	EECS1015 - Introduction to Computer Programming

• Deep Learning Engineer at Eagle Eye Networks (April 2024 - March 2025) - Toronto, Canada
  
  • Built multi-cue gun localization algorithm (pose estimation + action recognition + palm localization); boosted recall 13%.
• Deep Learning Engineer at UncannyVision / Eagle Eye Networks (Aug 2019 - July 2022) - Bangalore, India
  
  • Delivered production CV stack (Person Re-ID, detection/tracking, domain-adapted anomaly & background subtraction) and developed an Anchor Search Algorithm for finding dynamic anchors in one-stage object detectors, adding 3% mAP on in-house baselines.
  • Built a proprietary Site-Specific Training pipeline that raised detector mAP 8% with 500 annotated images and designed a semi-supervised clothing-color classification model which increased the accuracy by 5%.
  • Engineered edge-ready vision modules: 70 FPS Grid-Temporal-Median background subtraction and an unsupervised reconstruction-prediction LSTM trajectory-anomaly detector (98.4% precision, 99.2% recall).
  • Redifining One-Shot Object Detector for Two-Class Problem
  • Low-Compute - High FPS BGS Algorithm for Low End Devices
  • Site Specific Training Tool
  • Vehicle Trajectory Anomaly Detection
  • Person Retrieval and Re-identification
  • Unsupervised Domain Adaptation and Semi Supervised Learning
• Founder and Editor at VisionWizard (Mar 2020 - April 2022) - Ahmedabad, India
  • Mentored in publishing more than 25 deep learning research blogs (140k reads).
  • Publication Link
• Machine Learning Engineer Intern at UncannyVision / Eagle Eye Networks (Jan 2019 - April 2019) - Bangalore, India
  • Pruned and quantized models for edge deployment, delivering 3x speed and 2.5x smaller footprints.
  • Implemented TensorRT from scratch for a further 1.2% mAP and 1.5x throughput boost.
  • Network Pruning of Deep CNNs
  • INT8 Quantization of Object Detectors

    Ablation Studies and Results can be obtained from my BTech Report: Link

---

Patents

• US Patent US20230072641A1 - Image Processing and Automatic Learning on Low Complexity Edge Apparatus and Methods of Operation

  One of the inventors of this patent focusing on edge-based image processing and automatic learning systems.

---

Open-source Contributions

• FreeMatch Algorithm - Unofficial implementation of FreeMatch algorithm for semi-supervised (SSL) image classification.

• Cross Domain Adaptive Clustering - Contributor of Cross Domain Adaptive Clustering algorithm in Dassl.pytorch for unsupervised domain adaptation (UDA) image classification.

• Multi Radius Deep SVDD - Authored Multi Radius Deep SVDD algorithm, an improvement on Deep One Class (DOC) Anomaly Detection.

---

Projects

• Intelligent Traffic Light Control System

  Created a prototype for automating the traffic light timers by doing dynamic timer set-stop transitions at the crossroads for Ahmedabad. Density calculation was done through two different approaches viz Deep Learning: TFNet YOLO Detection and Computer Vision/ Image Processing: Foreground extraction.

• Classification of Gender from Tweets and their Profile Description

  Gender classification from Tweets and their profile description through Natural Language Processing for formatting the tweets and C 4.5 Decision for featurization and Shallow Neural Networks for classification of the gender from the formatted tweets. We reached 84% accuracy on the test set

• Thread Management System

  We developed a thread management library to create, schedule, and kill threads(Operating System). All these operations were done on the kernel level with the use of Linux system calls. The library had similar functionalities as pThread in Linux.

• Saliency Map Generation from Images

  Extraction of saliency maps from two different approaches i.e., Supervised and Semi-Supervised. The supervised approach follows discriminative features integration with superpixel image segmentation followed by finding saliency scores with Random Forest Regression. The semi-Supervised approach follows Autoencoder based approach with an attention layer to find the saliency maps from images.