GenAI: AI for Coding

Interned in the AI Infra org. I worked on improving the data IO transparency of AI pipelines at Meta.AI pipelines can be complex, reading data from multiple sources and writing to multiple sinks. My project helped automatically identify such sources and sinks with 80% accuracy. I leveraged a combination of Static Analysis with Machine Learning to identify the callgraph of a pipeline and to identify IO functions using a Large Language Model based classifier.I designed the system to be scalable from ground-up given the large number of dependencies involved in such pipelines.

Helped teach CS 411 - Database Systems for two semesters. Mentored teams on their projects, reviewed paper critiques, designed homework and exam questions, conducted office hours, and helped with automating some workflows.

I worked with Dr. Venkat Padmanabhan on two problems:1. Improving cost efficiency of Microsoft's WAN provisioning: Microsoft spends a huge sum on their WAN network every year, and we worked out a way to reduce the amount of required network capacity by a significant percentage. This technique, when applied, would result in saving hundreds of millions of dollars annually. This work has resulted in a publication at NSDI 2021, a premier conference in Networking and Systems.I led the work on developing and improving tooling around their WAN simulator and network provisioning tools, in order to make them work for our analysis. This involved working on a very large codebase with little guidance, and I reduced simulation run time from 10s of hours to 10s of minutes. I also contributed equally to the core technique that we came up with.2. Automatically deducing latency tolerance of users based on existing telemetry, through the use of natural experiments as opposed to A/B testing. This work involved working extensively with 100s of TBs of data using Microsoft's big data ecosystem, and performing data analysis with python. Our paper describing this work has been published at ACM IMC 2021.

I worked on improving IBM's blockchain platform - Hyperledger Fabric. I, along with my mentor Dr. Senthilnathan Natarajan developed 3 key techniques that would result in significant performance, scalability and cost efficiency improvements in the platform. We developed a way of pipelining transaction validation, enabling sharding, and enabling auto-scaling for Hyperledger Fabric. Our evaluation showed that our improved version gave 3.7x higher throughput and provided original peak throughput at 50% to 87% lower cost!This work involved the intellectual effort of developing the key ideas of the 3 techniques, and a heavy implementation effort for the realization of these techniques - my work added over 15,000 lines of Go code. Apart from this, I also developed the tooling around running 1000s of experiments over a cluster of nodes.Our paper, titled "Scaling Hyperledger Fabric using Pipelined Execution and Sparse Peers", describing this work has been accepted at ACM SoCC 2021.

My responsibilities include making and maintaining websites for college, and webapps for college festivals Pragyan and Festember.As a part of Delta, I also take active part in hackathons and other programming contests.I teach at workshops conducting by Delta.

Worked on benchmarking and optimizing Hyperledger Fabric v1.0. Improved the overall throughput by 16x from 140tps to 2250tps. Published a paper describing the same at MASCOTS '18. It won the best paper award there.

I worked on an Echo-driven Tech Conference, where Amazon Echo devices use voice recognition to act as help desks and MoC’s.I Built a fully serverless backend on AWS to power FireOS apps, the Echo devices and awebsite. I wrote integration tests and handled deployment and Ops of the whole system.The system was successfully deployed and used at Amazon’s India Tech Conference ’16.