Sign In

Mukesh Agarwal

Mukesh Agarwal is a Partner in the AI for Customer Experience division at Ernst & Young (EY). He has a significant background in leveraging artificial intelligence to enhance customer interactions and experiences. His current role involves leading initiatives that focus on developing AI-driven solutions to improve customer engagement and satisfaction.

Professional Background

  • Current Position: Partner, AI for Customer Experience at EY.
  • Focus Areas: Mukesh specializes in generative AI and customer experience strategies, actively contributing to the development of AI-based platforms that measure and personalize customer interactions. He has been involved with the SMX platform, which is designed for customer experience measurement and personalization.3

Education and Skills

Mukesh Agarwal has a strong educational background, including affiliations with prestigious institutions like the MIT Media Lab, which reflects his commitment to innovation in technology.4 His skills encompass a blend of consulting, technology implementation, and strategic leadership.

Online Presence

He is active on LinkedIn under the username mukeshagarwalonlinkedin, where he shares insights related to AI, customer experience, and industry trends. His posts often emphasize the importance of staying ahead in technology to provide exceptional customer service.12

Mukesh's professional journey showcases his dedication to integrating advanced technologies into business practices, making him a notable figure in the field of AI-driven customer experience solutions.

Highlights

Aug 23 · linkedin

Learning from the LLM trenches around optimizing the LLM operations -

#1 - Optimizing the LLM output is a science problem, not an engineering problem. The engineering approach is to "define the problem - build a solution - troubleshoot the solution - bug-fix" it. This may not work for GenAI optimization. A science approach is to "define the problem - set up experiments - run those experiments - systematically measure the output - pick the winning experiment." and rinse and repeat. So, it's more akin to data science than data engineering. [prompt engineering is the wrong term] #2— Models are trained on a corpus of data from the open web—text, code, images, videos, etc. Interact with models in the format they are trained on. If you invent a new way to present some information alien to the model, they may interpret it but likely not as well as if the information was presented in the format they are trained on. #3- Reduce "cognitive load" on a model in a single LLM operation. Models are trained on a large amount of data, but their reasoning capabilities are still of the 3rd grader. Break down a complex "task" into many simple "asks." #4 -If you are building a complex workflow app, don't try to solve every problem through LLM. If the problem can be solved outside the LLM operation, do that. Rely on LLMs to solve the core analysis and reasoning problem. #5 - While models like GPT4o are improving support for structured output, they do so at the expense of accuracy and quality. They use constrained decoding, which means if you ask in, say, <answer>..</answer> format, the model will pick up the "<answer>" tag from the format you define and run the next word completion. This constraints the vector space it looks up for the next word. [This is a super simple explanation]. You can try to get the answer in the raw format and use another prompt to structure it in the format you want if you must use LLM for the structured output.

In short, keep the interaction/instructions for LLM simple. It can better comprehend what it is trained on, and its reasoning capabilities are still primitive.

#genai #llms

Related Questions

What are some key projects Mukesh Agarwal has led at EY?
How has Mukesh Agarwal contributed to the development of the SMX platform?
What is Mukesh Agarwal's background before joining EY?
How does Mukesh Agarwal integrate AI into customer experience strategies?
What are Mukesh Agarwal's interests outside of work?
Mukesh Agarwal
Mukesh Agarwal, photo 1
Mukesh Agarwal, photo 2
Add to my network

Location

Bengaluru, Karnataka, India