Description:

In current times, Building a generative AI application sounds easy, right? Just follow a few tutorials, and you'll have a generative AI/ RAG app or AI agent up and running in no time. But what happens when you deploy it in production? Suddenly, none of those tutorials seem to work as promised. The harsh reality is that most online resources focus on building proof-of-concept models, not robust, production-ready applications. As a result, many developers struggle to overcome the challenges of deploying generative AI models in real-world scenarios. In this talk, I'll share my own experiences and lessons learned from building and deploying generative AI applications. We'll dive into the common challenges and pitfalls that you won't find in abundance of online tutorials, and explore practical solutions and best practices to overcome them.

This talk will focus on

• Evaluations: How to effectively evaluate your AI application and what metrics really matter.
• Prompting Strategies: Techniques for crafting effective prompts that gets you the desired responses from your large language model, and how to avoid common pitfalls
• Structured Output: Strategies for generating structured output that's usable in real-world applications, and how we can minimise the JsonDecodeErrors. This is one of the most important use of LLMs as downstream applications always accepts structured information to process it.
• To FineTune it or not: One of the thought of every business / developer. When should we finetune or when should we limit to RAG or both?
• Practical Advice: Tips and tricks from my own experience, including how to debug issues, optimize performance, and ensure model reliability.

The PLAN

This talk will be application focused, by the end we will be having a personal AI assistant responsible for generating us a balanced workout routine based on past history

• (First 5 minutes) The Current Hype
• (Second 5 minutes) The Reality of AI in production
• (Next 10 minutes) Best Practices and Practical Advice

Having made many AI applications both for clients and my personal use-cases, I have got stuck a countless times. As soon as I started using the applications as a daily driver. I have always found some of the issues, the pattern is, 99% of the times the issues are same and everyone will face it. I with my year of a journey in building practical applications on top of LLM will share my lessons and what worked and what didn't.

Prerequisites:

There are not much pre-requisites, but to have fun in talk, it's good to have the following understanding - Foundation knowledge of programming - Basics of APIs / JSON - Have interacted with a chat based LLM (example ChatGPT)

Video URL:

https://youtu.be/M_wsQGNl2HM

Content URLs:

This is an active area of my work as an indie-maker. The repository is a work in progress

• GitHub Repo for the project

I am working on slides currently, will share once first cut is available.

Speaker Info:

Aditya is a passionate Data Scientist and Architect at TaskHuman, responsible for unlocking the full potential of data to deliver a rich and personalized experience for users. With a strong foundation in machine learning, natural language processing, and data engineering, he designs and develops AI-powered applications that drive business value. Currently, I'm working on various projects that leverage AI to solve real-world problems, such as personalized workout planning, deterministic agent planning and custom chatbots with Retrieval Augmentation Generation. I strive to stay at the forefront of AI and data science, continuously learning and experimenting with new technologies to push the boundaries of what's possible to create a meaningful impact.

Speaker Links:

• My Corner on Web
  • 𝕏
  • LinkedIn