Oct 2020 – Present · 5 yrs 5 mos · Lausanne, Vaud, Switzerland

• Sequential prediction of music with free polyphonic texture.
• The problem of predicting polyphonic music is difficult from the perspective of its huge prediction space. The existing solution tries to solve this problem using either “black box” deep machine learning architectures or make far-fetched musical assumptions. The deep learning models fail completely in communicating a musical process in doing so. An interpretable model for polyphonic music will not only help composers in better understanding styles in music but will also aid in musicological corpus studies and music transcription. We solve this problem by defining a set of interpretable musical features, learning a policy that uses these features to predict polyphonic music. We solve the problem of large prediction space by having a nested Markov process running underneath the main polyphonic prediction problem which incrementally constructs a timeslice by adding pitches sequentially. Finally, we discover new and more nuanced musical features by using thePULSEframework to expand the current set of features. We train the model on 2 corpora, J.S. Bach Chorales and Mozart Sonatas. The model successfully predicted 91.7% on the implicit test for chorales and68.1% for sonatas. The model was able to learn the rules of parsimonious voice leading and constructing harmonious chords along with staying away from dissonances.

Jul 2020 – Sep 2020 · 2 mos · Remote

• Physics-based reasoning problem is a challenging and open research field. Several works have explored the idea of learning the environment dynamics. In this project, we explore the ideas of Model-based reinforcement learning in the field of physical reasoning puzzles. We learned the physical model of the environment from the sequential images collected from the PHYRE framework. We used a convolutional-RNN to learn the latent embedding of the sequences of images. We performed experiments on 2500 physical reasoning puzzles in the framework. Our novel contribution in this work was the network design which was able to capture the underlying model dynamics. This approach also allows us to plan in the latent state-space without the need of decoding back into the observation space of pixels.

Jan 2019 – Jul 2019 · 6 mos · Greater Hyderabad Area

• I joined Delhivery as a senior developer to help a team of 23 members restructure and rewrite their warehouse management system. In my short time there, In addition to structuring their new WMS, I also identified a critical bug in their current system to reduce their loss due to unidentifiable items by 10%.

Jan 2016 – Jun 2018 · 2 yrs 5 mos · Hyderabad Area, India

• Demand-side platform, Programmatic ads, Real-time bidding.
• Initial member of the engineering team started with a mission of bringing the right job closer to the user, through recruitment ads on the internet. Our mission is to show the user right ads from our clients' recruitment ads. I took part in coming up with product vision, designing the architecture. I am part of the team that built a highly scalable product to serve our product vision.
• Part of the job is to understand and implement machine learning algorithms and data science techniques, to power our smart machine optimized job distribution engine
• Technologies:
• Scala
• MongoDB
• AngularJs/Js
• S3, Loadbalancers, SQS, Redis, Lambdas

Aug 2014 – Nov 2015 · 1 yr 3 mos · Greater Hyderabad Area

• Part of the Salesforce File sync Team, A native application for Windows and Mac to synchronize files across multiple platforms.