Generative AI on AWS - Amazon Bedrock, RAG & KIRO [2026]

Explore the evolution of generative AI and how artificial intelligence, machine learning, and deep learning interrelate, with examples like self-driving cars and recommendation engines.

Explore how artificial neural networks mimic the brain with input, hidden, and output layers; understand weights, bias, activation function, and backpropagation that drive deep learning.

Explore the fundamentals of generative AI, including how foundation models generate text, images, and other data, with examples from OpenAI's GPT and AWS Nova models on Bedrock.

Discover how a prompt is tokenized into tokens and token IDs, processed in a context window to predict the next word, then de-tokenized into a completion with inference and streaming.

Foundation models are deep neural networks trained on massive unlabeled internet data, enabling text generation, summarization, Q&A, and code generation, with tokens, parameters, and temperature shaping outputs.

Discover Amazon Bedrock, managed serverless service exposing base models via API from Amazon and third parties; pick a model ID, set prompts, and call Bedrock via console, cli, or sdk.

Explore Amazon Bedrock architecture, including the service and escrow accounts, runtime inference, and model endpoint routing. See how private link and VPC interface endpoints secure requests and data.

Explore the Amazon Bedrock console walkthrough to learn the model catalog, serverless and marketplace models, API keys, testing, and inference options including cross-region, batch, and provisioned throughput.

Explore Amazon Bedrock's multi-agent framework, flows, knowledge bases, and data automation to build generative ai workflows with drag-and-drop simplicity, harness rag and automated reasoning, guardrails, and prompt management.

Explore how inference parameters shape foundation model behavior on Amazon Bedrock, detailing randomness and diversity controls—temperature, top K, top P—and length controls such as max tokens and stop sequences.

Build a production-ready generative AI powered equipment SME assistant to reduce wind turbine downtime by summarizing field incident logs with Amazon Bedrock foundation models, boosting productivity and revenue.

Design a PoC-to-production workflow that streams wind turbine logs from a streamlet UI through API gateway to Lambda, using Amazon Bedrock Nova Pro with system prompts and inference parameters.

Build and deploy an AWS Lambda function to send the Equipet SME logs to Amazon Bedrock, invoke the Nova Pro foundation model for summarization, and return results via API Gateway.

Learn how to invoke Amazon Bedrock from Lambda by crafting the request with the model id, a body containing user input and inference parameters, and optional guardrails and latency settings.

Implement bedrock invocation in lambda using a system prompt to summarize logs, configure inference parameters, and return the final response to aws api gateway.

Build and secure a REST API with AWS API Gateway connected to a Lambda function, configure resources and mapping templates, and deploy to a prod stage for end-to-end testing.

Test your API externally with Postman by sending a POST request to the API gateway, using the prod stage and a JSON body to receive a three-line Lambda summary.

Set up observability for generative AI on AWS by enabling Bedrock logging and CloudWatch metrics. Configure log groups, retention, data types (text, images, embeddings), and alarms with SNS.

Apply Amazon Bedrock Guardrails to protect generative AI apps with content filters, denied topics, word filters, sensitive information filters, and contextual grounding checks to mitigate hallucinations and toxicity.

Enable bedrock guardrails in the equipment SME assistant by configuring guardrail id and version, deploying code, and validating through test events and api gateway to ensure toxic prompts are blocked.

Learn how to select foundation models for your use case by evaluating modality, context window, guardrails, cost, latency, and regional availability, using Bedrock and Nova Pro as examples.

Learn Bedrock model evaluation strategies for a GenAI equipment SME app, using programmatic, llm-as-judge, and human evaluations with prompt datasets, jsonl prompts, reference responses, and metrics.

Optimize foundation model outputs through prompt engineering by crafting contextual prompts, clear task and output specifications, and exploring zero-shot to few-shot prompting plus retrieval augmented generation.

Learn how model distillation transfers knowledge from a large teacher model like Nova Pro to a smaller Nova Lite using prompts, achieving near large-model accuracy at lower latency and cost.

Explore hands-on model distillation on AWS Bedrock by creating a distillation job that trains a student model from a larger teacher model using synthetic data and S3 datasets.

Utilize supervised fine tuning with labeled prompt–response data on Amazon Bedrock, triggering a SageMaker training job to tailor a model's style and format for tasks like summarization and Q&A.

Extend model knowledge through continued pre-training on proprietary unlabeled data, such as medical journals, to improve domain-specific physician note summarization within Amazon Bedrock using SageMaker-backed training.

Explore a mental model for foundation model adaptation strategies, including prompt engineering, retrieval augmented generation, distillation, fine-tuning, and continued pre-training. Evaluate trade-offs in speed, cost, accuracy, and enterprise data use.

Follow a five-step roadmap to become a generative and agentic AI architect on AWS, mastering genai foundations, bedrock fundamentals, rag, and bedrock agent core deployment.

Trace AWS's AI-powered software development evolution from code whisperer to Amazon Q developer and finally Kiro, a spec-driven approach that generates requirements, designs, and code with iterative refinement.

discover how kiro, AWS's agentic coding service, drives spec driven development from prompts to requirements, design, code, and tests with a clear, traceable workflow.

Learn to install and set up Kiro with IAM Identity Center on AWS, using Frankfurt or US East region, creating users, sending invitations, and choosing a Kiro plan with credits.

This lecture demonstrates using Kiro to transform a vehicle diagnostic feature into a set of requirements in a markdown file (requirements.md), capturing AI-driven diagnostic conversations and interaction history.

Generate design.md and task.md files with Kiro from the requirements, analyze requirements for consistency, and log key design decisions while creating a task list and MVP plan.

Kiro integrates with external tools within enterprise via Model Context Protocol to push user stories and design documents to Jira and Confluence, translating natural language queries into tool calls.

Learn to integrate Kiro with an MCP server using Jira and Confluence, configure MCP, push user stories from requirements.md to Jira, and publish a design document to Confluence.

Implement steering files in FleetMate by configuring DynamoDB-backed AP layer with API gateway, Bedrock, and Lambda, integrating product.md, tech.md, and structure.md for contextual responses.

Execute all mandatory tasks from the requirements, build FleetMate end-to-end, and deploy the cloud infrastructure, frontend, and backend on AWS using SAM, CloudFormation, and Kiro.

Automate repetitive development tasks with Kiro hooks by triggering agent prompts or shell commands on IDE events, updating documentation and tests to speed delivery and improve quality.

Set up hooks in Kiro to automate tasks triggered by events, including a security pre-commit scanner, and define actions like running commands or asking Kiro, while updating the agent context.

Discover kiro powers, pre-packaged bundles of documentation, steering files, and MCP configurations, enabling context-aware loading and standardized outputs. See how power.md semantic matches activate the right MCP tools, reducing tokens.

Explore hands-on use of Kiro powers to install and configure AWS observability tools. Monitor CloudWatch logs, metrics, alarms, and application signals across us-east regions.

Explore Kiro model options and AWS recommendations for selecting the right model for each task, starting with auto, and using Opus for complex problems or Haiku for quick iterations.

Build a chat bot with Amazon Bedrock and Llama two, using Lang chain memory and the conversation chain, with Streamlit for the front end and Bedrock prompts for context.

Configure the chatbot environment with VS Code, Python, AWS CLI, and an IAM role, then install Boto3, LangChain, Streamlit, and PyYAML for Amazon Bedrock and AWS generative AI.

Builds a chat backend by wiring a conversation memory buffer, an Amazon Bedrock LM, and a conversation chain, then invokes the model to generate replies.

Build a chatbot frontend with Streamlit, importing the backend and memory, customizing the UI, and running the app to connect OpenAI or Amazon Bedrock.

Demonstrates an end-to-end chatbot workflow using Amazon Bedrock, Llama 2, and the cloud foundation model via Streamlit, with a frontend interface and Lang Chain integration.

Enable Bedrock model invocation logging to CloudWatch, configure log groups and storage options (CloudWatch, S3, or both), and capture request/response metadata, tokens, and latency for all model invocations.

Build a rac hr q&a use case with Amazon Bedrock and Anthropic, using LangChain to orchestrate integration with a policy pdf for leave details.

Architect a HR Q&A app using retrieval, augmented, and generation, built with a data ingestion workflow that loads pdfs, creates vector embeddings, and stores them in FAISS for Bedrock-powered responses.

Set up your chat bot environment in vscode via Anaconda navigator, then install flask, SQL alchemy, py pdf, and the Facebook AI Similarity search vector store to start coding.

Build an HR Q&A app using retrieval augmented generation by loading a policy PDF from a URL with a pdf document loader and LangChain, then test data loading.

Split pdf text into paragraphs, lines, and characters using a recursive character text splitter to form 100-character chunks with 10 overlap for embeddings. This prepares data for a vector database.

Create vector embeddings and store them in a vector store, perform similarity search, and send the context to Amazon Bedrock's Claude model to generate HR answers.

Build the rack frontend with Streamlit and connect it to the rack backend to create an end-to-end hr q&a app with rag.