Working on GPU Kernels (Triton / TLX / Cutlass / CuteDSL), New Hardware Optimization (e.g. Blackwell), Training Efficiency for large scale recommendation models.

Selected work:

• Built one of the first Jagged flash attention kernel widely deployed at Meta

• Key contributor to Generalized Dot Product Attention

• Blackwell optimization for Meta GEM model

• Co-created and led Ad's kernel library

Leading ML@Drishti. We build cutting edge computer vision models for the factory floor.

We are solving various exciting problems for productionizing state-of-the-art computer vision models.

These include:

• Exploring promising research directions such as semi-supervised learning, few-shot learning, meta learning, spatio temporal action localization, vision transformers, etc.

• Building a Self Serve AI and MLOps platform to easily train and deploy models, continuously monitor and retrain models on mistakes for maintaining daily accuracy.

You can read/watch more about our work here:

* https://analyticsindiamag.com/how-drishti-empowers-deep-learning-in-manufacturing/

* https://www.nvidia.com/en-us/on-demand/session/gtcfall21-a31579/

As a founding engineer, I worked on various problems across Drishti:

* I led the action recognition research from the beginning. We use state of the art deep learning models to do real time action recognition on the factory floor

* I developed models that can end to end learn large molecular actions from data - reducing the need of rule engines (heuristics)

* I designed and built Drishti's ML platform for continuous training and evaluation

* I designed and built Drishti Data Platform which collects all the NN generated data and builds views on top of it (Kafka, MongoDB, Bigtable)

* Led the Kubernetization of services at Drishti - helped teams move to CI/CD

Worked / Working on the following projects:

1. Deep Learning based Visual Similarity system for Flipkart. Arxiv Paper: (https://arxiv.org/abs/1703.02344)

2. Neural Network Retail Modeling (For recommendations)

Working on Social Sentiment Analysis, Review Mining, Email Categorization, and other NLP related problems. My approach to solving these problems is based on Deep Learning, a field which I'm very excited about.

Gave a talk on "Deep Learning for NLP"​ at HasGeek's FIfth Elephant conference. https://fifthelephant.talkfunnel.com/2015/64-deep-learning-for-natural-language-processing