Building AI Apps: From Concept to Production

Imagine a world where AI-powered applications seamlessly enhance our daily lives. 🌟 In this exhilarating talk, Pratik will guide you through the fascinating realm of Large Language Models(LLMs), demystifying how they empower developers to create cutting-edge AI projects.
Large Language Models (LLMs) are at the forefront of this revolution, transforming how users search, interact, and generate content. But wait, there’s more! Recent stacks and toolkits, like Retrieval Augmented Generation (RAG), allow users to build chatbots and other applications using LLMs on their private data. 🤖🔍
However, setting up a naive RAG stack is a breeze—productionizing it is the real challenge. Fear not! Pratik will delve into core techniques for evaluating and enhancing your retrieval systems, ensuring top-notch performance for your RAG applications. 📈

Transcript

1. Building AI Apps From Concept to Production
2. So, you are building an AI App 🥳 What are the problems you’re going to face? 🤔
3. PokeBot Your own Pokemon Analyst to answer all pokemon related queries!
4. • Easy to build, Hard to scale • Fine-Tuning vs Prompting vs RAG • The Ambiguity of Natural Languages • Inconsistency in User Experience • Prompt • Cost monitoring • Observability and monitoring Problems I Faced 😢
5. Problem hai?
6. Problem hai? Aage solution hai!
7. Pratik Parmar CTO, Stealth Startup Ex. Zyte, Intel, yellow.ai
8. RAG vs Fine-Tuning vs Prompting RAG Fine-Tuning Prompting Student has a textbook of that subject. In the exam, student can refer from the textbook to answer the question. Teaching students on specific subject and then ask questions related to that subject in an exam. Asking students to write answers in the specific format only. For i.e. MCQ, Short note, Long answer.
9. RAG vs Fine-Tuning vs Prompting - Solution! RAG Fine-Tuning Prompting Data in evolving. More data required, Higher accuracy. Less example, quick and easy to start.
10. The Ambiguity of Natural Languages Expected output format Essay score: {score} / 10
12. The Ambiguity of Natural Languages
13. Inconsistency in User Experience Expected output format Essay score: {score} / 10
15. Inconsistency in User Experience Set temperature = 0 Deterministic, repetitive More creative, Random Temperature 1 0
16. Inconsistency in User Experience
17. Inconsistency in User Experience
18. Prompt Engineering Few shot Add 3+3: 6 Add 5+5: 10 Add 2+2: One shot Add 3+3: 6 Add 2+2: Zero shot Add 2+2:
19. Prompt Engineering: Few Shot
20. Prompt Evaluation: Overfitting
21. Prompt Evaluation Check if your model understands example given in the prompt
22. Prompt Versioning
23. Prompt Optimization • Chain of Thought ⛓ • Multiple Output 🔢 • Break prompt into smaller tasks 🔪
24. Cost 💰
25. Observability and Monitoring 🔍
26. Observability and Monitoring 🔍
32. PokeBot is Ready to Roll! 🎉
33. What are you going to build with AI now? Share your ideas with me on Linkedin https://linkedin.com/in/pratikparmar1