Salman Maqbool

- Working closely with mobile developers for continuous improvement of server-side and mobile-side deep learning models and pipelines - Trained, converted, optimized, and deployed text and multimodal Transformer models for iOS and Android - Defined relevant metrics and designed and implemented an end-to-end evaluation and MLOps system to keep track of those

- Worked closely with mobile developers to develop data acquisition and client-side 3D model library apps (launched on App Store and Play Store) - Architected, developed and optimized a modular backend to receive 3D data from mobile apps and run object, scene, and human body reconstruction pipelines to generate 3D models - Introduced Agile and GitOps to the team for better development practices and developer productivity - My work directly contributed to the company getting a contract from KT Corporation

- Researched novel approaches for large-scale yet fine-grained visual classification, incorporating the Destruction and Construction Learning (DCL) training paradigm to improve our fine-grained classification accuracy - Defined different evaluation metrics and implemented those so that we can continuously monitor our performance - Enhanced code readability and performance by redesigning and implementing it as GStreamer modules, optimizing deep learning models using DeepStream and TensorRT

- Developed and deployed an LSTM research paper classifier with an accuracy of 91 %, reducing curation time and effort - Analyzed discrepancies in our reference matcher. Improved matching performance using ElasticSearch

- Worked with Dr. Faisal Shafait and Dr. Ahmad Salman on fish detection and classification in challenging underwater environments using both conventional and deep learning techniques - Achieved state of the art 84.5 mean Average Precision on the Fish4Knowledge dataset using Gaussian Mixture Models augmented with Pixel-Wise Posteriors - Designed and manufactured apparatus for underwater video capture in freshwater reservoirs in Pakistan for data collection

- Worked on simulation and track reconstruction of beam telescopes for the ATLAS and CLIC experiments - Configured and simulated the FE-I4 telescope in the Allpix-squared simulation framework - Extended Allpix-squared to output data in the right format for track reconstruction - Extended Proteus track reconstruction framework to account for the effects of multiple scattering by incorporating the General Broken Lines algorithm in the framework

Trained and deployed Deep Learning networks for fine-grained vehicle detection, classification, tracking, and counting of cars in Pakistan

- Setup software and projects, and provided relevant IT support to designers and engineers. - Programmed and deployed Excel worksheets on Servers which helped our department determine the active users of different software products; organizing the data and saving our department’s time.

Machine Learning, Computer Vision, Artificial Intelligence, Robotics, Deep Learning

Mechanical Engineering, Instrumentation and data acquisition, Control systems