Past
Projects
Listing some of the projects I worked upon in the past recent years. These projects have helped me to develop intuition in problem solving and how to quickly onboard a new problem statement.
As an applied AI engineer, I have worked at the intersection of research domain & industry projects commercialization. This has made me better at reading research papers as well as putting them into production.
I carry a wide exposure to industry level challenges and this experience makes me a good fit to tackle real-world problem statements.
Plotting graph of equations and their intersection
An app to show how to plot the graph of a given function equation. Extending it further to plot the points that lie on the intersections of the given functions.
Key learnings:
Text-to-Speech model for Hindi language (Flipkart)
The aim of this project was to generate audio speech in Hindi language for the given text sentences. I started on this project while I was working with Flipkart.
Key learnings:
-
Unsupervised generative models
-
Autoregressive models
-
Flow (& inverse Flow) based models
-
Papers I've read - Link
MNIST-Muddle (Personal project)
The aim of this project was to get a practical exposure of how Latent Domain works in deep learning. This is a simple project which tries to generate poorly written hand-digits by interpolating between the nearest cluster.
Got practical exposure to:
-
Auto Encoders
-
Latent domain
-
PyTorch hosting using Streamlit
-
Project Demo - Link
C++ decoder for Speech recognition engine (Flipkart)
Worked on the decoder module of the ASR pipeline (Automated Speech Recognition).
Key responsibilities:
-
Implemented new features into production.
-
Handled __ lines of C++ code base.
-
Improved latency & memory consumption.
-
Blog post 1 - Intro to CTC Loss
-
Blog post 2 - DP based prefix beam search PPT
Single Image HDR reconstruction & enhancement(Samsung)
Camera sensors typically have low dynamic range which often leads to over-exposure or under-exposure of regions in outdoor scenes. This project aimed at enhancing the overall image by improving such regions.
Key learnings:
-
Un-paired images DL training
-
Cyclic loss, encoder-decoder
-
HDR, tone mapping, linear vs non-linear images, histogram
Demosaicing using CNNs (Samsung)
A camera sensor captures single channel image where each pixel contains only 1 color information. The process of interpolating the remaining 2 colors via neighboring pixels to construct a full colored image is called Demosaicing. We used a deep learning model to approach this problem.
Key learning:
