Experience

Machine Learning Engineer - III Feb, 2021 - Present

InFoCusp Innovations Private Limited, Ahmedabad, India.

Client: (Confidential Research Project)

  • Large Language Models for Code Synthesis
    • Developing transformer-based models for program representations and using them for code completion, code repair, and code translation.
    • Developed fully automated and hybrid ML pipelines for data engineering, modeling, evaluation, and deployment that can work on both client's private infra and GCP. Some of the frameworks I used were TensorFlow, Kubeflow, KFserving, and Apache Beam.
  • Computer vision
    • The goal was to create a single virtualized view of the electricity system through an aerial view imaginary using an image segmentation model.
    • I explored various state-of-the-art image segmentation models such as PSP-NET, ViT, etc., and combined the best of all in a single model to achieve production-grade results.
    • I also developed a novel metric for the evaluation of the model's performance as traditional segmentation metrics failed to provide helpful indications of failure cases. While the new metric helped filter failure cases easily which was then improved through data augmentation techniques.
    • GRAD-CAM and its variant were used to locate features in the image when the model failed to produce the correct output. This analysis helped improve the quality of the training and evaluation dataset.

Client: Innovyze, An Autodesk company.

  • At Innovyze, I helped Process Engineers to optimize chemical consumption in the water treatment plant by analyzing historical sensor data and developing predictive models to automate the processes. My primary tasks were Data analysis, Feature engineering, and Modelling.

Software Development Engineer - II (AI/ML) May, 2018 - Feb, 2021

Matrix Comsec R&D, Vadodara, India.

Matrix ComSec is a leader in Security and Telecom solutions for modern businesses and enterprises with 1M+ customers in 50+ countries. I was responsible for developing and delivering DL algorithms in SDK form. My major contributions are listed below,

  • Face Recognition (FR) and Face Detection (FD)
    • Developed FD algorithm to detect multiple faces in the image with a minimum face size of 30px.
    • Developed FR algorithm with 99.85 % accuracy on the LFW benchmark.
    • Improved existing FR algorithm to identify people wearing a mask on the face with only 5% accuracy drop wrt to full-face model. It helped employees to mark attendance without removing their masks during the coronavirus pandemic
    • Developed a Face Mask Detection model to identify whether employees are wearing masks or not when marking attendance to improve their safety.
    • Developed CNN for a single RGB image-based Passive Face Antispoofing model to prevent fake attendance marking using a mobile phone or printed photo.
    • Delivered all these features in a single FR SDK for RPi3/4, CPU, GPU, and Android devices with a maximum latency of 800 ms on the slowest hardware
  • Automated License Plate Recognition (ALPR)
    • Developed a real-time Licence Plate Detection model which achieved 99%+ recall and 93%+ precision on the internal Indian vehicle test dataset
    • Contributed to the development of the CNN model for Licence Plate Recognition using CTC Loss. We achieved 87% accuracy on challenging the Indian test datasets outperforming other commercial solutions by at least 15% and 98% accuracy on the OpenALPR benchmark
    • Developed semi-supervised data annotation tool which helped us collect a labeled dataset of 30K+ images and incrementally improved the model's accuracy
    • Designed ALPR SDK architecture for improving throughput by batching simultaneous requests into single inference. A myriad of inference engines such as TensorFlow, TensorRT, OpenVINO, TensorflowLite, etc. can be integrated based on hardware without modifying 95% of the codebase.
  • Seven Segment Display Number Recognition
    • To handle a variety of color digital displays, a self-calibration algorithm was designed to run once to extract display-specific properties using KNN and an inference algorithm was designed to use these properties for recognizing digits in real-time using SVM.
    • The algorithm was integrated into existing weighbridge vehicle management software for automating data entry of the vehicle's weight from the digital display through the camera. It reduced manual human intervention by 95% and introduced transparency in the whole process.

Deep Learning Research Intern Jan, 2018 - May, 2018

Space Applications Centre (SAC) - ISRO, Ahmedabad, Gujarat, India

SAC being the major R&D center of ISRO designs and develops the optical and microwave sensors for the satellites, signal and image processing software, GIS software, and many applications for the Earth Observation (EO) program of ISRO. As a Research Intern, I developed DL techniques for accurate crop classification using Hyper Spectral Satellite images (425 channels per pixel). My primary contributions were as listed below

  • Developed ANN-based dimensionality reduction technique which retained useful features and worked better than PCA.
  • Developed virtual data augmentation technique for overcoming the issue of insufficient ground truth data of different crops for training.
  • Provided detailed comparative analysis of ANN, CNN, and SVM. Designed highly accurate Parallel-CNNs architecture for classifying nearly inseparable classes based on interclass separability analysis.
  • Developed Python library based on my work so that other scientists can use these techniques on a variety of other satellite images.

Machine Learning Summer Intern May, 2017 - June, 2017

Wolfsoft Pvt. Ltd., Vadodara, Gujarat, India.

During my internship I worked on a food review app similar to Zomato and my contributions were as listed below,

  • Developed a Food Recommendation API based on a Collaborative Filtering algorithm.
  • Developed a Food Search API. Normalized and stored posting list of Indian food names in the Trie data structure for faster and more accurate search. Added spelling correction using the Levenshtein distance algorithm which achieved higher recall compared to SQL's search.