Emile Baizel I Building GenAI applications with Amazon Bedrock I Agentic Ethereum 2025

my name is Emil basil I'm a blockchain Solutions architect at AWS I've been here for over five years at this point uh prior to here I was uh doing solidity development over at consensus and then have been um mostly working as a blockchain architect uh since joining AWS and thank you eth Global for this opportunity to uh participate in this in this hackathon um at AWS we've been doing a lot of work over the last couple of years um in uh in in generative Ai and um more recently in uh agentic workflows and so everything that that that we're going to be talking about here and everything that this hackathon is about um is really close to what we are doing at AWS and we're always excited to see what what developers build with the tools and the products and services which we which we develop So today we're going to be talking about Amazon bedrock and Bedrock agents uh but to start off I want to give a highle overview of how we view generative Ai and how we're incorporating it across our stack so here we have a sort of three- layer stack of uh different different levels of uh integration with generative AI so at the top you have user phasing applications uh you have Amazon Q developer for example this is a coding assistant it plugs into VSS code other idees um which you may be using uh in the middle we have Amazon Bedrock which is what we're going to be talking about today and then at the bottom is for machine learning engineer so if you're writing if you're building a model from from the ground up you would use Amazon Sage maker uh a lot a broad choice of um different Hardware uh which you want to run that on but that's for a different day today we're going to be talking about that that middle piece with Amazon bedrock so Amazon Bedrock is a fully managed service that offers a choice of foundation models from leading AI companies such as AI 21 Amazon anthropic meta and many other ones uh with a broad set of capabilities to build many different types of generative AI applications and a big focus on really simplifying development uh while maintaining privacy and security so Amazon Bedrock is fully managed it's serverless so you don't have to worry about any of the infrastructure you access it via apis you access it via sdks um so uh very easy to get started and and build with there and with Amazon Bedrock you can easily experiment with a variety of these models and you can customize them privately with your data or using external data uh either via fine-tuning or uh retrieval augmented generation or rag uh and you can create managed agents for creating um executing complex business tasks such as booking travel or processing insurance claims or other things like that and all of that can be done without writing any code um I assume this being a hackathon people will want to get get into the code but just to give you an idea of um of the building blocks which you are providing uh you could go into the AWS console click a few things enter a few things and have an agent that is running that can uh crawl a website and answer questions about it and things like that cool get this uh get my PowerPoint set up there we go okay next slide so we speak to a lot of customers at AWS and what we hear a lot is that what the most important feature to them of bedrock is how easy it is to experiment with different Foundation models and having a range of models to experiment with uh customers are moving fast in this space and generative AI itself is evolving very very quickly and there's new options and Innovations happening almost daily and just last week we saw uh uh the market turned on its head with uh deep seek R1 and um and we moved very quickly and now deep seek is available within bedrock I'll get to that um towards the end uh so the space is moving very quickly everyone knows this and uh Bedrock is really a great way to easily test out different models and uh and experiment without having to um do a lot of heavy lifting to to change that again everything's via apis and so you uh so if you're changing a model for example you're basically just changing the label of the model which you want to um invoke cool I do want to touch on two uh two families of models the first is anthropic claw models uh these have been out for over a year now I think they're almost a year I should say um and these are familiar to many people uh Cloud 3.5 Sonet uh really raises the bar in like industry intelligence and in benchmarks for graduate level reasoning and you got claw 3 Opus which is the most ENT model and uh three Sonet and three Haiku with three Haiku being the fastest and most cost effective and that's really good for sort of high volume use cases so it's really dependent on the type of application which you are looking to build again testing with any of these is U is is very straightforward to switch between them and uh prior to reinvent uh which for those of you who don't know it's our uh annual conference that takes takes place right after Thanksgiving uh we had the Amazon Titan model uh family of large language models and at reinvent we announced the Amazon Nova Foundation models and these are new generative generation these uh state-of-the-art Foundation models uh that deliver Frontier intelligence and industry-leading price performance and so Nova includes uh a couple types of models so there are understanding models which are these ones on the left and these accept text and image and video as input and they generate text as output and on the right side you have creative content generation models so these accept text and image input and they generate image or video output and Nova models are available only on Amazon bedrock and the five that are available today are Nova micro which is a text only model that delivers the lowest latency responses at very low cost uh you have Nova light that's a lower cost multimodal model that is also very fast and is great for lightweight tasks and then you have Nova Pro that's a highly capable multimodal model uh with a good combination of accuracy speed and cost and on the right side you've got our creative content um FMS Foundation models so canvas is an image generator using natural language prompts and then Nova real is a video generator uh which takes prompts and creates realistic videos from that and in the middle is Premiere which is not launched yet but it's uh it will be our most capable multimodal uh model um Foundation model and that's for complex reasoning and also to be used as a best teacher for distilling custom models so if you're trying to U distill down to a smaller model you can use this one as the starting point from that and Premiere is coming soon and as I mentioned earlier uh deep seek this slide was added uh uh on Friday or over the weekend because that that announcement just just came out but as you can see on the right there that's the AWS console and now deep seek is one of the generation models uh one of the foundation models which you can use within Bedrock cool so we just touched on inference and selecting different llms for your inference now I want to switch gears and talk about agents so Amazon Bedrock agents are foundational building blocks for customers and developers to create their own agentic applications so as we just saw you can choose from a variety of foundation models to drive the intelligence of the agent and you can configure the agent with many different features such as memory retention knowledge bases for uh retrieval augmented generation uh guard rails for um protecting the user and um and other features as well and developers provide these agents with additional tools and functionality so that the agents can actually go and uh achieve their tasks and um if you've ever worked with AWS Lambda functions that's essentially what a tool is here is you equip the agent with this capability and you tell it when to invoke this uh this tool for example if you want to create a calculator you just tell the agent use this this this thing here and it's a Lambda function that basically just does simple math or something like that and an important thing to understand here too is um with agents for Bedrock you get visibility into the Chain of Thought of the agent so when you're debugging and you're trying to figure out why why your agent isn't doing what you think it should be doing you can see step by step with all the traces of everything that that it's thinking through and why it reached a certain decision which which helps immensely when you're when you're building and and and debugging and uh recently we announced uh Bedrock agents support hierarchical multi-agent collaboration so you have what's known as a supervisor agent and this agent can have its own intelligence and its own uh tools but it can also be aware of other agents that can uh invoke that that can be invoked for specific tasks and you can now have orchestration from this sort of supervisor higher level agent um and these agents can all be used across different uh orchestration as well so you can have one uh sort of sub agent that's being used by many different uh supervisor agents and agents uh plan and execute these tasks using uh data which you've provided either from your own from your own private data set or from public data it can really be driven uh from many different ways and as I mentioned earlier you can you can start creating agents just from the AWS console so if you are interested in um just seeing what that looks like and setting up a Lambda function um you click you click through there and then you get a Lambda function that's a basically Hollow World dummy function and then you can easily extend it to uh to um do the actual work which you need it uh to do cool so the way agents work is they essentially break down and orchestrate the task so when when a user gives it a question or a task the agent thinks through everything which it needs to do and then it figures out and creates a plan for how it's going to accomplish that task and to help it accomplish this task you can allow it to securely access and retrieve data from from your companies or projects uh data sources or anywhere else and this would be part of um retrieval augmented generation to to to help your agent um sort of become more specific to your application um and it can take actions on your behalf so it can invoke API calls on on your behalf and that's where we get into the uh action groups and the the the tooling which you equip your your agent with and then we've got the Chain of Thought trace and ability um and the ability to modify the agent prompts as it's going through that Trace um so we have a default orchestration um uh template for for agents but with with a lot of these Services we have a sort of uh best best default but a lot of a lot of uh a lot of what I'm covering here you you can customize it to to Really um tailor it for the application which you're building so if you're interested in trying out bedrock and multi-agent collaboration um here's the first of a few QR codes which I'm going to uh share this is for the multi-agent U which we which we just released recently um and some examples which you can see are uh planning a trip where you have different agents that specialize in different aspects of planning a trip uh building a marketing uh strategy and other things like that and these are all available on on GitHub cool and the last thing I want to show is uh as of this weekend we got our PR merged into uh Eliza OS and so now your Eliza OS agent can use Amazon Bedrock for its intelligence and I wrote this PR so if you try it and it doesn't work you know exactly who to shake your fist at um but please let me know and I would be happy to um help you debug any issues or answer any questions which you have around this all right so mov moving on from Agents and I want to talk about retrieval augmented generation so as as many of you probably know what an llm when it's created is uh Frozen in time at that time where it was created so it doesn't know anything that that happened afterwards right and so that's where retrieval augmented generation really helps because you can supply data to the agent or to the uh during inference um for the agent uh to to to provide context around uh something that the agent may not may not know about either because it happened after or um it's private data so it was never available for it to be uh indexed and that's where uh Bedrock knowledge bases come in so rag really uh allows the agent to get data from your company data sources and then it enriches the prompt with that data to deliver more relevant and accurate responses and knowledge bases is a fully managed rag capability that allows you to customize your foundation model responses with that contextual and relevant data and uh where knowledge bases really help developers is it it automates the entire end to-end rag workflow so this includes uh data ingestion data retrieval and augmenting the prompt and so this eliminates the need for writing custom code to integrate these data sources to manage the queries and again it's uh something you can set up just from the console where you click a few buttons set set your data source and um and you have it go so for example you can you can have an S3 bucket with you know hundreds or thousands of PDF files and maybe you want to make these available for your agent so that it can have more knowledge about whatever is in those in those PDF files um you set those in S3 you go through the console you select that as your data source and then you set a few other options and then that be now becomes uh available to your agent so your agent can now query that that data and I'll get into how that works in a second cool so when you're starting with knowledge bases you start by selecting what um what what Vector store you want to use so our default is Amazon open search serverless uh but knowledge bases natively support uh postgress with PG Vector via Amazon Aurora also supports pine cone and a couple of other ones or if you have an existing database you can link it to one of those so you have a lot of options when it comes to um what what Vector database you want to use uh we already talked about S3 as a data source you can also use a a WebCrawler so if you want to crawl websites and set that on a iteration of uh you know crawling it every five minutes or something else like that it'll pull that out and it'll pull all that data in index it and all of that is managed for you uh you can also connect it with uh Native connectors which uh knowledge bases has and that's typically more for Enterprise data but I just want to mention it here uh things like Salesforce or if you're using Confluence or um SharePoint for example you can connect those automatically and have those be available from uh from your knowledge base so once you've defined what your vector store is once you've defined what your data source is then you need to define the chunking strategy and this is where the data the the data itself so if we go back to that PDF example um that PDF then gets chunked into many different chunks and added to the vector database and there's a bit of trial and error with figuring out the right chunking strategy and chunking is basically how big each how big how many um uh pieces you want to separate the the the document into how much overlap you want to have within the chunks because remember what we're trying to do here really is uh make it so that when a user queries or asks something related to this the knowledge base will be able to retrieve from the vector database something that is relevant and so we have some smart defaults but you can play around with it and see if if if you're getting the results you want and um and you can get even more advanced where not just um setting the chunking size but you can actually go in and override the chunking itself and an example here is if you want to add metadata to each chunk right so that so that that could help with with um more semantic search right so instead of just indexing um the content of the PDF you can actually Supply additional metadata about that about that PDF and so that'll make it much more likely that it gets that it gets pulled in all that to say it's very easy to get this set up but if you want to get in the weeds you can also get in the weeds with a lot of this stuff cool um couple more things we continue to add new features uh to knowledge bases um yeah was there a question no all right there there is a question um from Mii um he asked can I return the logits uh probability for a prompt um I'm going to ask to save questions till the end because I just want to have the presentation be unbroken but I would love to answer that question at the end and I forgot to mention this at the top thank you Vara for um speaking up there uh following this call we're going to have uh some office hours so that'll be a great time to ask uh questions like that or any other questions you have so we continue to add new features to knowledge bases uh recently we introduced Amazon Bedrock data Automation and that really helps you extract transform and generate structured data from multimodal content and so all of that gets automated for you and made available into your knowledge base and we also introduced graph rag which as the name suggests is it makes your data available in a graph database and so what that helps you do is start linking relationships uh for your entities across data sources so instead of just having one one view of of your data here and one view there you can start linking them together uh within a graph database uh the last thing I want to touch on uh not related to generative AI but because it's a ethereum hackathon it's a blockchain hackathon I want to share our public um blockchain data this this is uh free data it's available on um AWS uh public data sets and it's a continuously updated set of data across three different blockchain or seven blockchain networks and uh and and the way to think about this is the data gets um gets gets parsed from the blockchain it gets put into an S3 bucket and then you can access that data from your application and the QR code here is a project that myself and Simon from my team worked on last summer where uh it's essentially a text tosql querying of this data and and so the way that that works is you get your prompt from the user your llm we we um instructed the llm and made it uh aware of the of the of the schema that that is uh that that all the data um abides by and so then the when the user prompt comes in the large language model can say Okay so the user wants to know how many ethereum transactions there were yesterday and it knows which uh which which columns in the database map to that that type of question and it constructs a SQL query and then it goes and queries that so if you're building something where um uh where you want your agent to be aware of historic data um this could be a great way uh for for you to bring that that information in there cool and the last thing I'm going to leave you with is three uh three three QR codes here and these are uh for helping you get started so on the left is our documentation for Amazon Bedrock it's pretty broad I mean as we saw here we've got llm inference we've got knowledge bases we've got agents and so uh but that's a great place to start uh if you like uh YouTube videos we have a lot of YouTube videos there uh that's the one in the middle it's a stepbystep tutorial and if you're looking for open- Source samples of uh examples of what what uh projects have been built with uh Bedrock that's the that's the QR code on the right um it's a pretty broad set of applications um and and demos which have been built there and with that I want to say thank you uh for for for listening and and watching and uh yeah let's turn it over to some questions if we have some