Natural Resources
Canada

• Building classical neural networks with parameter constraints to classify image data for comparative analysis with quantum neural networks.
• Conducting detailed comparisons to explore the interplay and advantages of classical and quantum neural networks.

Ford Motor Company

• Implemented an automated bug ticket triaging system using word embeddings and a machine learning classifier, achieving 80% accuracy in routing tickets based on titles and descriptions, saving 15 hours of total developer time per week.
• Analyzed over 2 billion log lines, extracting insights such as high-memory processes and low-impact crashes, leading to fixes that reduced crashes by 20% and memory usage by 10%.
• Automated configuration validations and detection of common issues, integrating these solutions with Jira, which reduced the number of tickets requiring manual triaging by 50%.
• Parsed over 1 billion log lines to build a classification dataset of over 2000 data points for machine learning, extracting key features based on developers' manual analysis processes.
• Applied machine learning classifiers to the dataset, effectively handling imbalanced data using ADASYN, and achieved an F1 score of 82%.

Avoma

• Fine-tuned transformer-based embeddings, using bi-encoder and cross-encoder architectures to build a semantic similarity model which achieved an accuracy of 88% on meeting transcript data.
• Leveraged LLMs, along with RAG, particularly GPT-3.5 and GPT-4, for various summarization, rephrasing, question-answering, and sentence classification tasks, to deploy 6 new features, including transcript summarization and automated notetaking.
• Implemented image processing and OCR techniques to identify the current speaker from virtual meeting interfaces (Zoom, Google Meet, MS Teams), improving the accuracy of the current speaker recognition system from 67% to 88%.
• Developed a feature using clustering on over 100,000 meetings to rank top FAQs and reveal key themes; its deployment increased average CSAT score from 7 to 8.
• Involved in system and data platform design for deployment of these models, using Django, PostgreSQL and AWS EC2 and ECS to build end-to-end Machine Learning inference and training pipelines for real-time sentence classification and transcription.
• Optimized inference speed of Tensorflow models, using the gRPC protocol and AVX acceleration achieving a 30% decrease in latency.
• Set-up and maintained automated CI/CD pipelines, using Tensorflow serving, Docker and Github actions, automating and reducing deployment times from approximately 10 minutes to 2 minutes.

Helpshift

• Developed advanced machine learning solutions to automate end-user issue resolution in customer service.
• Conducted research, scoping, benchmarking, and deployment of cutting-edge NLP algorithms, including user message classification and language detection.
• Engineered scalable microservice inference pipelines for machine learning models using Django, Flask, and FastAPI, which supported over 1 million requests per day.
• Diagnosed and resolved critical bugs, optimizing underperforming models for clients, by using oversampling or adjusting class weights on imbalanced datasets, boosting classification accuracy from 70% to 90%.

Segmind

• Contributed to the development of CRAL, a library used for abstraction of well known deep learning architectures for Computer Vision.
• Worked on the addition of well known deep learning architectures for computer vision into the library.
• Intensively involved in implementation, integration and testing of object detection models.
• Achieved mAP scores of more than 0.6 on standard benchmark datasets for all the integrated object detection models.

• Developed a novel Generative Adversarial Network architecture to increase the resolution of images.
• The input to the network is a low resolution image, which is upscaled natively by the network.
• The model is capable of upscaling input image by 4x the original resolution.
• The metrics obtained from our DPSRGAN are better than the previously proposed SRGAN, with a MOS of 3.91 out of 5 and a PSNR of 32.24.

• Developed a pipeline for analysis of tree census data using data of Pune, India.
• Introduced a novel metric, the Flora Biodiversity Index (FBI), to quantify the diversity of trees in a region.
• Drew insights from the data to determine uniformity of tree cover, areas deficient in trees and areas having a lower biodiversity.
• Our pipeline will be useful for cities to analyse their current green cover and work on making it better.

Languages

Python

C/C++

MATLAB

Java

Clojure

SQL

Javascript

Libraries

NumPy

Pandas

OpenCV

Matplotlib

Seaborn

Frameworks

TensorFlow

PyTorch

Scikit-Learn

Django

Flask

FastAPI

Selenium

Developer Tools

Git

Docker

Databases

MySQL

PostgreSQL

MongoDB

Cloud Technologies

AWS (S3, EC2, Lambda, ECS/ECR)

Tensorflow Serving