Evolve to AI

#personalinsights  #AIagent  #LLM

2024 was the year of AI agents, right? Yes! (if you’ll check in LinkedIn posts and some newsletters) However, production implementations tell a different story. Current AI agent solutions predominantly rely on basic approaches like ReAct, tool use, and routing mechanisms - or are essentially LLMs with minimal agency capabilities.
This session introduces a sophisticated alternative for building AI agents through dynamic planning and smart execution. I will review existing methods as well as this novel method, demonstrating its practical advantages through a live coding demonstration. Throughout the session we will build together an example application using the existing methods and the novel approach introduced here, comparing key metrics including output quality and development velocity.

#livecoding  #architecture  #AIagent  #howto

AI is one of the hottest trends currently. Everyone talks about generative AI, and thinking about how to add AI capabilities and use LLMs to extend the business capabilities of their servers. In this talk, Tamar will focus on a specific capability in the AI world - function calling. We will dive into how function calling - how to implement it, and how it can add to the user experience and business value of the application while writing an application that use function calling. At the end of this talk, you will have the knowledge of how to combine function calling in your server and on which cases to do so.

#modelevaluation  #performance  #LLM

Large Language Models (LLMs) are pivotal in driving user-centric AI solutions, but bias within these models can undermine their effectiveness and profitability. In this session, we will share concrete examples of GenAI chatbots making incorrect assumptions that can hurt user experience and erode profits. We will then showcase an open-source dataset that is a gold standard among Big Tech for assessing LLM fairness and accuracy. Lastly, we will provide developers with tools to balance model sensitivity and accuracy effectively, enabling them to tailor models to meet diverse business use cases.

#moltimodel  #AIagent  #voiceAI

Discover how multimodal AI can reshape user engagement, unlock new revenue streams, and spark groundbreaking applications across industries. In this session, Shai will show you how to build real-time voice and vision agents with Gemini 2.0, revealing how to integrate audio, text, and video while adding personality and practical tools for tangible impact.

You’ll learn how to go beyond traditional chatbots by orchestrating voice, video, and text in scalable, production-grade architectures. By the end, you’ll be ready to create innovative, AI-driven solutions that captivate users and transform the way they interact with your product.

#performance  #finetuning  #SLM's

Step into the fascinating world of small language models (SLMs) and discover how they rival their larger counterparts with greater efficiency and adaptability. This session unveils cutting-edge techniques for domain adaptation through fine-tuning, proven strategies for measuring success, and an inspiring real-world financial use case. Learn how SLMs are transforming specialized applications and shaping the future of AI innovation.

#text2image  #devops  #cloud 

How do you bridge the gap between cutting-edge generative AI models and production-ready features? Text-to-image and video models are evolving rapidly, deploying them effectively in production can be straightforward with the right framework.

What if you could harness these powerful models without getting tangled in complex infrastructure? In this talk, we'll explore how the ComfyUI framework transforms this vision into reality. Drawing from my experience at LTX Studio, a product by Lightricks, I'll demonstrate how we've used this approach to build groundbreaking features like generative fill and identity preservation solutions that are already serving users in production.

By the end, you'll understand not just the 'what' but the 'how' of bringing sophisticated vision features to life. You'll walk away with concrete strategies for transforming the latest AI breakthroughs into practical, production-ready features.

#cybersecurity  #AIagent  #LLM

Security professionals face the challenge of extracting actionable insights from the overwhelming flow of information on platforms like X.com (formerly Twitter). This session showcases an AI agent that streamline real-time threat intelligence from social platforms, demoed through X.com. By using an LLM-powered search engine for user discovery, the X API for live content retrieval, and a fine-tuned LLM for content analysis, the system generates precise, actionable threat intelligence feeds.  
 
The demonstration will delve into the system's technical architecture, exploring the advantages and limitations of various LLM models, including the trade-offs between large general-purpose models and smaller fine-tuned ones. It will also provide guidance on selecting a model, preparing it for training, and tailoring it to specific research needs—in this case, processing threat intelligence feeds from X.com.

#livecoding  #personalinsights  #moltiagent  #evaluation 

What is the advantage of having a team of agents instead of just one agent doing everything? Imagine a top-notch developer who excels at coding, product design, and architectural planning—could they single-handedly manage an entire project?

What if we could clone this developer? Would a team of specialists collaborating be more effective than individuals working in isolation?

To explore this, we'll observe agents in action as they respond to our request to create an application, we will compare it to how a "traditinal" team and a single agent tackle the same development task. Next, we'll experiment with dividing the team into specialized domain verticals to evaluate the impact on the results.

Finally, we'll analyze how applying patterns and lifecycle approaches—similar to those used with human teams—influences the performance and quality of agent teams.

Could insights from AI teams inspire the way we structure our real-life teams?

#moltiagent  #devtools  #gaming 

Discover how a simple text prompt can transform into a fully realized 3D game world. In this session, we’ll explore how a text-to-games engine leverages chain-of-thought reasoning and multi-agent systems to automatically generate code, assets, and emergent gameplay, revolutionizing game development and unleashing unprecedented creative possibilities.

#industryinsights  #productivity  #innovation

Join me for an illuminating talk on harnessing generative AI to catalyze transformational change. I’ll dive into our groundbreaking 2024 GenAI initiatives that are driving cross-organizational productivity and sparking product innovation. Through detailed case studies and real-world examples, discover how these innovations are reshaping industries and delivering impactful results.

#livecoding  #generativeUI  #frontend

Step into the future of web applications as we explore how to leverage Large Language Models (LLMs) to generate interactive UI components that deliver intuitive user experiences.

Through practical demonstrations, you'll discover how to approach multi-step workflows while effectively managing context and state using React's server-side rendering capabilities.

By the end of this talk, you'll be equipped with the tools and techniques to transform your innovative ideas into production-ready applications powered by generative UI.

Get ready to be inspired!.

#automation  #genAIapps  #LLM

In this session, We'll examine a real-life use case in the insurance industry, demonstrating how summarization, sentiment analysis, information extraction, and object detection can save up to 80% of the time spent on claim processing while reducing human errors. we'll explore the integration of Predictive and GenAI models in modern applications.

#casestudy  #search  #hands-on

Writing effective search queries is difficult, especially when users have limited input. Traditional search engines rely on keywords, but Generative AI and semantic search allow users to express themselves more freely and receive better results. At Fiverr, we’ve combined these technologies in our Fiverr NEO chat-based search engine to deliver more relevant results.
In this session, I'll share how we built a hybrid search system that blends keyword and semantic search, using Generative AI to improve query understanding and continuously optimize results with real-time feedback. We'll code together a chat-based search engine from scratch, see how to evaluate its results' relevance, monitor it in production and scale it up using VectorDBs and traditional information retrieval approaches.
You’ll leave with practical insights on how to build your own AI-driven search engine, leveraging semantic search and Generative AI to enhance user engagement and improve search relevance.

#productivity  #devtools  #testing

AI code generators excel in benchmarks but often fall short in the real world. Why haven’t they revolutionized development? Because they generate code snippets that lack context, duplicate functionality, require manual integration and can introduce errors.

This talk presents a new approach to code generation that reuses existing code, follows company standards, tests for impact, and generates fully executable components ready for integration.

#pronpedengineering  #moltimodel  #devtools

This presentation explores the transition from prompt engineering to flow engineering as a necessary evolution in AI's cognitive capabilities. While early LLMs like GPT-4 operate largely through fast, intuitive responses akin to System 1 thinking, newer models such as OpenAI’s o1 aim to embody aspects of System 2 thinking—characterized by deeper, more deliberate reasoning processes.

We will explore how flow engineering frameworks can enhance AI performance by guiding models through iterative reasoning, validation, and test-driven workflows. This structured approach pushes LLMs beyond surface-level problem-solving, fostering more thoughtful, strategic decision-making. The presentation will show how these advancements improve coding performance on complex tasks, moving AI closer to robust and autonomous problem-solving systems.

Key Takeaways:

Understanding of test-driven flow engineering to help LLMs approach System 2 thinking
Assessing how well models like o1 & o3 tackle complex coding tasks and reasoning capabilities
The next generation of intelligent software development will be multi-agentic AI solutions capable of tackling complex challenges with logic, reasoning, and deliberate problem-solving.

 #livecoding  #performance  #LLM

Speculative decoding is a neat algorithm for LLM inference acceleration. Join us as we uncover how it works - in a live coding session!
The idea is relatively simple: Instead of generating tokens using a big model - which costs you time and money - you can use a smaller one to generate the tokens more efficiently, and then verify them using the big model. The result? Faster and cheaper text generation, while not compromising on the quality. In fact, there's a mathematical proof that shows that you end up with a text distribution identical to the text distribution of the big model.
Join us to learn how you can implement Speculative Decoding in 20 minutes - from zero to hero!