Strutfit is New Zealand based SaaS provider for footwear size recommendation to ecommerce businesses like Hoka, VivoBareFoot and Bobux. Unlike the indirect and time consuming foot measurement techniques used by its competitors, Strufit has employed advanced AI and mobile photogrammetry for accurate and repeatable shoe size recommendation. Our goal was to enhance the overall foot length accuracy and repeatability and to incorporate foot width measurement required by footwear manufacturers.Contributions:Research and Development Proof of concepts for Foot Width measurement (Metatarsal & aligned bounding width)Foot measurement technique using Plastic Cards Transition from traditional Machine Learning techniques to powerful Deep LearningIncorporated Foot Rotation and image rejection due to high alignment anglesDesigned and deployed training and inference pipelines on AWS and GCP along with MLoPMemory efficient Object detectionSubpixel semantic segmentation under variable lighting conditions Optimized and fine tuned models for higher accuracy, GPU memory and loading timeDesigned custom ELT pipelines for data cleaning and segregation (in S3 buckets)Microservices architectureSmartphone Camera calibration and using already estimated camera matrices and distortion coefsAccuracy enhancement using Transfer Learning and Data augmentation using custom entity relationship in the imagesCode review and standard Agile practicesHR training on Deep learning, image segmentation, object detection, photogrammetry, AWS/GCP, and image labelling tools like LabelBox, SuperAnnotate etc.Technologies used in this journey:Languages: Python, C/C++, C# for backend and R&DFrameworks: Tensorflow, FastAPI, Flask, MLFlow AI/Analysis: Semantic segmentation using Deeplab, Detectron, Yolo, Unet, Opencv, Keras, PandasCloud/Deployment: Azure CICD, AWS ECR, ECS, SageMaker, EC2, S3, Lambda, Cloudwatch, Docker, TerraformVersionning/Agile/Management: Git, Slack, Asana, Notion, Jira

Street Planning System: Neurontix USA Analytics dashboard for Traffic planning over CCTV infrastructure containing wide angle cameras on a 3 lanes street. The planning system contains a track of every vehicle passing through each lane, its type, parking spot density and stop duration in long term and delivery lanes. The system also identifies the delivery vans and brand advertisement on trucks. All the traffic stats were recorded into a database and traffic analytics were available on a ReactJS Application along with timestamped images and videos.ContributionUtilized transfer learning for improving the publicly available vehicle detection model using our custom dataset. Enhanced vehicle detection performance under low lighting and hazy conditions from 72% to 88%. Another challenge was to calibrate the cameras and to preserve the vehicle appearance in all area of the image. Due to low order estimation in standard libraries the edges of the scene were prone to maximum distortion. Developed a higher order non-linear curve fitting for the fish eye correction which ramped up the overall vehicle detection under severely distorted conditions to 95%.Technology StackPython, ReactJS, Postgres DatabaseProduce Counting System: HelloFresh New ZealandDesigned the vegetable and fruit detection and counting system for Hellofresh which can count up to 13 classes with the accuracy of 99% over a conveyor running at its normal speed. The imaging system was designed using realsense cameras, LED lights, and industrial graded lasers. Nvidia AGX board runs the image analysis containing DeepSort as real time object detector and tracker. 3D reconstruction was also applied to enhance the accuracy and to suppress the false positives. The system produces outputs to an HMI which could also be used to provide various inputs to the control system & image analysis script.Technology Stack:Python, DeepSort, opencv, RealSense, Laser Triangulation, LED control, AGX Jettson, HMI

Part of Scott's team that envisions state-of-the-art industrial robots being used in G.E, Fisher & Paykel and other major production lines. The goal is to improve the overall throughput of our robots using real-time computer vision.Responsibilities:R&D in Process automationRequirement analyses and surveysProject planning and client liaising Solution ArchitectureProjects: Realtime optical Coordinate Measurement Machine Design for Door line inspectionMilkTech: Autonomous Cow Milking System using KuKa Robot, ToF camera Lambcut: Lamb cutting system for Silver Fern FarmScott's Industry 4.0: Research report

Been a part of NZ top Laser Tag providers (DeltaStrike) as a system engineer.

An an IT/Cloud infrastructure manager; I have been closely working with the directors for introducing and implementing innovative ideas to expand the business over cloud. Designed and developed their Ecommerce system with integration of modern and secure payment gateways and ERP system. Another achievement was to enable the Printer Genie feature on the main website that could detect the installed printers on user's network. By allowing this feature, users dont need to know the Printer models for searching their consumables(toners, drums, cartridges, cables, drivers etc). During the journey, we also moved away from classical DoS based Inventory Management to an improved GUI based integrated system.After moving out of Palmerston North, I had been working remotely for Inkpost until Nov 2017.

Worked with Prof Bob Hodgson on Classifynder/Veritexa project designed by Massey's research team. I have been involved with the optimization of pollen data classification using various morphological features from its Z-Stack. The system received NZ Excellence Award in 2014.http://www.veritaxa.com/

Mainly involved as a system engineer in a US based aircraft simulator designing company that was running various projects for Turkish Air Force, Pakistan Air Force and Royal Saudi Arabian Air Force. Been working on Matlab, C/C++ and assembly language for realtime imagery, instrument and flight control systems.