Increased throughput, decreased memory footprint, and enabled parallel scaling of the distributed batched inference service

Designed and implemented asynchronous LLM APIs / bots which allowed multiple users to have ChatGPT- like group conversations with Abacus LLMs in communication platforms such as Slack and Microsoft Teams

Created PySpark-based incremental datasets from database services for ETL into product

Built periodic full execution of cloud-hosted Jupyter notebooks for data science using scripting inside Kubernetes deployments

Implemented functionality to create / update custom ML models using customer-provided python code

Developed Cloudburst-based system which generates warm starts for ML pipelines

Added memory multiplexed GPUs / scheduling algorithms to Cloudburst serverless project

Deployed pre-trained ML models on the Cloudburst serverless framework to provide autoscaling resource allocation for inference functions

Built a space efficient algorithm for asynchronously aggregating and reporting keys across distributed server nodes in Cloudburst with high and low frequencies of access.

Worked on Ground project: www.ground-context.org with graduate student Vikram Sreekanti, in Professor Joseph Hellerstein's group

Added support for non-primitive datatype inputs for Ground

Created Ground implementation using Git repositories and functionality to construct a time-complexity efficient RESTful API for accessing Ground objects

Implemented a Python-based metric scraping framework for analyzing and storing timeseries data for AthenaX

Analyzed the resource usage efficiency of different types of AnthenaX pipelines