Download Updated PDF Resume Here
Education
University at Buffalo, SUNY, NY, USA Jan 2022 - Present
Masters in Computer Science and Engineering (GPA 3.83/4)
Coursework: Machine Learning, Computer Vision & Image Processing, Deep Learning for Biometrics, Information Retrieval, Natural Language Processing (NLP), Distributed Systems
University of Mumbai, India June 2018
Bachelors in Computer Engineering (GPA 8.17/10)
Coursework: Artificial Intelligence, Digital Signal Processing, Algorithms, Data Structures, Operating Systems
**Experience **( ~ 4 years)
SUNY Research Foundation - A2IL (Graduate Research Assistant) March 2022 - Present_
- Chart Infographic Data Extraction - Extracting data from line chart images
- Improved accuracy of data extraction by 12% in multi-line chart images by modeling it as a Graph Min-Cost flow problem, analogous to Multi-Object Tracking and solved using Linear Programming.
- Isolated background grid lines in chart images, and segmented chart lines of different styles (solid, dashed, etc.) using Local Binary Patterns and Color Histogram features.
- Developed a pipeline for generating synthetic degradations to digital chart images using Image Processing to mimic noises in archived document images.
Publication: LineFormer - Rethinking Line Chart Data Extraction as Instance Segmentation (Under Review) International Conference on Document Analysis and Recognition (ICDAR) 2023
Karza Technologies_ (Data Scientist - Computer Vision) April 2020 - January 2022 (1.8 years)_
- Robust Document Text Recognition & Script Classification:
- Reduced Text Recognition Character Error Rate by 11%, by making the model invariant to different image distortions using synthetic feature supervision & spatial transformer networks.
- Improved accuracy of the Script Classification model by 8% by using an auxiliary text recognition loss, also resulting in better generalization on out-of-domain text images.
- Document Quality Estimation:
- Developed a system to evaluate OCR-suitability of an image based on several parameters such as Blurriness, Noise, Resolution etc. using Deep Learning and Image Processing.
- ML Deployment, Optimization and Security:
- Increased throughput and reduced inference time by 10-15% for several Deep Learning models serving more than 10K requests/day in production using **TensorRT **inference engine, and other techniques like model quantization, batch inference, etc.
- Developed scripts for securing ML models using obfuscation** **for deployment on mobile apps.
_ _
Artivatic Data Labs (Computer Vision Engineer) August 2019 - March 2020 (8 months)_
- Developed a template-based Handwritten-Forms data extraction pipeline involving Image Registration, **Denoising, Word Recognition, **all optimized for mobile deployment.
Barclays Bank_ (Graduate Analyst) July 2018 - August 2019 (1 year)_
- Contributed to the development of a Robotic Process Automation framework, by identifying desktop GUI elements (buttons, scrollbars, text boxes) using Image Processing and OCR.
Academic Projects
American Sign Language (ASL) Recognition April 2022
- Modeled a Deep CNN for classification of ASL words from acoustic signals captured as spectrogram images and obtained a test accuracy of 93.5%
Pitchfork Music Rating Prediction March 2022
- Used music meta-data data for predicting pitchfork review score using Decision Tree Regression, exploring feature selection using Lasso on, and interpreting them to identify key factors influencing review score.
Volunteered Distributed Computing Architecture (VoDCA) Smart India Hackathon 2018
- Proposed a system for utilizing public compute nodes for scientific computation tasks like Aerial Image Stitching.
- Explored distributed execution of Image stitching pipeline, involving - Feature Extraction (SIFT, SURF), Matching, Homography & Image Blending.
Technologies
Languages: Python, Java, C/C++ | ML Toolkit: PyTorch, Tensorflow, Scikit-Learn, OpenCV, Numpy, Pandas
Deployment: TensorRT, AWS Neuron, TFLite, ONNX, Flask/Django, Git/GitHub
Accomplishments
- Recipient of 'RF Tuition Scholarship SEAS (Spring’23)' for contributions in research at A2IL.
- Core contributor in the Document AI team at Karza winning the award for NASSCOM AI Challengers 2021
- Delivered a well-received tech-talk on ‘Demystifying Deep Learning’, to an audience of over 100, including VPs, & Directors of corporate banking at Barclays.