Palli Padmini

About Palli Padmini

I contribute to end-to-end pipelines for building enterprise applications with low latency and optimization. I have expertise in various AI technologies, such as cognitive decision models, physics-inspired models, neural translation systems, generative AI, statistical models, federated learning, and LLM\'s. I also deliver directional approach across the technology stack and mentor the teams to achieve functionality in safe agile methodology.In addition to my industry experience, I am pursuing my Ph.D. in Speech Signal Processing at Amrita Vishwa Vidyapeetham, under the Meity (Visvesaraya) scheme. My research focuses on developing a real-time speech synthesis system for speech-disabled people using tongue orientation characteristics. I have published two papers in reputed journals and implemented a prototype device using flex and potentiometer sensors. My goal is to apply my data science and research skills to create innovative and impactful solutions for social and business problems.

Experience

1. Sr Associate - Projects

   Cognizant

   Mar 2024 — Present · IN

   Creating a pitch deck is traditionally time-consuming for analysts, often taking 20-25 hours per request. To accelerate this, we have designed a solution that leverages GenAI capabilities through Gemini, enabling faster deck creation by identifying optimal assets based on user inputs and prompts using RAG & LLM\'s. Developed a programmatic solution for bulk marking threads as Off-Topic or Abusive using the RPC studio with forum and thread IDs through a designated endpoint. Leveraged Stubby calls to trigger the RPC endpoint, enabling automated bulk closures for threads flagged as Off-Topic/Abusive, which increase the coverage for Google ads community by ~5%. Created an automated workflow to schedule monthly closures of flagged threads, ensuring efficient and timely moderation.Gecko embeddings and cosine similarity for open threads was implemented to find the best duplicate thread from Recommended threads which has accurately answered by PE\'s which helps to increase the coverage by ~20% by marking the open threads as dedupe. Automatically marked as dedupe with parent/recommended thread using RPC stubby call python. Handled over 10K+ open community threads monthly, aimed at reducing the manual effort required by product experts or specialists to review and respond to questions. Designed and implemented a GenAI architecture using RAG and Gemini models, integrating three key data sources: a camp skill tool, FAQ, and a help center website, to automatically find and deliver accurate answers to community thread questions, which aims to improve the coverage to ~85-90%.Utilized the Gemini 1.5 Flash GenAI model to classify Google Ads community threads, fine-tuning prompts to handle a range of questions, from generic information-seeking inquiries to complex troubleshooting scenarios. This enhanced the accuracy and efficiency of thread classification and response generation.

Education

• KSRM College of Engineering, Kadapa, Andhra Pradesh

  Master of Technology - MTech, Digital Electronics and Communication Systems

• Amrita Vishwa Vidyapeetham

  Ph.D, Speech signal processing

• Siddharth Institute of Engineering & Technology, Puttur

  Bachelor of Technology - BTech, Electrical and Electronics Engineering