-
Amsterdam, NetherlandsStaff Software Engineer, Intuit
Siva is a staff software engineer on the API Management Platform team at Intuit based in MountainView, CA. He works on building features for API Platform where developers can author, mock, explore and share APIs with other developers. During his free time, he enjoys going on hikes and runs a couple of marathons every year
Staff Software Engineer, Intuit
Rachit has spent over six years at Intuit, where his work has spanned from building platforms for monetization and AI powered conversation to enhancing user experiences in products like QuickBooks and TurboTax. Currently, he is part of an Applied AI team focusing on the innovative use of Generative AI to boost developer productivity through intelligent tools and methodologies, such as efficient GraphQL attribute discovery and dynamic query generation. Rachit looks forward to connecting with fellow innovators at this conference to exchange insights and discuss the evolving landscape of AI technologies and their applications in improving developer experiences.
Discovering GraphQL fields and generating queries is a tedious task for developers. They spend a considerable amount of time finding the appropriate fields in large schemas. To solve this problem at Intuit where we have a super graph consisting of millions of lines we implemented a framework that makes use of Generative AI to help developers with attribute discovery and query generation. The benefits of our approach include being able to work with large schemas without the hassle of going through the whole schema and requiring less back and forth communication between consumers and schema owners, which results in a huge boost in developer productivity. We created chunks of the schema and ingested them into a vector store, we then do a retrieval, dynamically build a minimal schema and perform RAG where the LLM is provided with the minimal schema and the user query. The LLM responds with either a list of discovered attributes or GraphQL query. This framework aims to achieve lower latency and less hallucinations by reducing the size of the schema sent to the LLM, this also results in lower costs and higher accuracy.