Generative AI with OpenAI Models on AWS
Contact us to book this course
Generative AI
On-Site, Virtual
1 day
This course teaches cloud architects and engineers to deploy OpenAI's GPT-OSS models using Amazon Bedrock's managed infrastructure. Participants learn to build production-grade generative AI applications that leverage GPT-OSS-20B and GPT-OSS-120B models with their Mixture-of-Experts architecture, all while maintaining enterprise security, governance, and cost controls through Bedrock's fully managed service. The course covers RAG implementation with knowledge bases, prompt engineering techniques specific to OpenAI models, fine-tuning and model customization, guardrails configuration, and architectural patterns for multi-region deployment with resilience and cost optimization across AWS regions.
Course Objectives
Through hands-on labs and real-world challenge scenarios, participants interact with applications that integrate S3, OpenSearch, Lambda, and CloudWatch with OpenAI models via the Bedrock APIs. The course includes practical guidance on model selection (GPT-OSS-20B vs. GPT-OSS-120B), understanding the OpenAI Harmony protocol for reasoning transparency, monitoring token consumption and latency, and implementing cost-effective architectural patterns, including cross-region inference, API Gateway integration, and PrivateLink networking. By the end of the course, participants understand how to design resilient, scalable, cost-optimized AI applications using OpenAI models on Bedrock with multi-region capabilities.
Who Should Attend
Solutions Architects, Cloud Architects, DevOps Engineers, Machine Learning Engineers, and Technical Managers building AI-powered applications on AWS using OpenAI models.
Prerequisites
- Understanding of cloud computing concepts and REST API interactions
- Programming experience in Python or JavaScript (recommended but not required)
- Familiarity with networking concepts (VPCs, subnets, routing)
- Basic familiarity with AWS console and core services (Amazon S3, AWS IAM, Amazon EC2, AWS Lambda)
Course outline
- Define generative AI and explain what foundation models are
- Understand the OpenAI model family: GPT-OSS-20B and GPT-OSS-120B
- Describe how OpenAI's models are available through Amazon Bedrock on AWS
- Explore Mixture-of-Experts (MoE) architecture and how it differs from traditional transformer models
- Demo: Invoke an OpenAI model via Bedrock using the Chat Completion API
- Walk through AWS's global footprint: regions, availability zones, data sovereignty, and latency considerations
- Understand Cross-Region Inference Profiles (future capability for OpenAI models)
- Explore disaster recovery architectures: application logic routing, Route53 + API Gateway + Lambda, Application Load Balancer + Global Accelerator
- Challenge Activity: Design a Bedrock-based deployment for a geography (e.g., APAC + Europe) to serve global users
- Define a knowledge base in the context of Bedrock, including data-to-embedding conversion and how retrieval-augmented generation (RAG) works
- Explore Amazon Bedrock Knowledge Bases as a fully managed RAG solution
- Configure data sources: S3, Web Crawler, Confluence, Salesforce, SharePoint, Custom
- Set up automated data sync with EventBridge
- Challenge Activity: Multi-region resilience for Customs Compliance Bot - design architecture showing Application, Knowledge Bases, and OpenAI Models across regions with a data consistency strategy
- Distinguish between fine-tuning (using labeled task-specific data), continued pre-training (using large, unlabeled domain data), and model distillation
- Explore Bedrock's Custom Model Import for OpenAI models (GPT-OSS architecture)
- Understand data preparation requirements: JSONL format and OpenAI Harmony format
- Highlight the advantages of using Bedrock for custom models: managed infrastructure, secure model lifecycle, global deployment flexibility, on-demand serverless inference
- Demo: Show model files in S3 bucket, import custom GPT-OSS model into Bedrock, use on-demand inference
- Challenge Activity: Calculate RAG vs. Fine-Tuning costs for LegalEagle AI contract drafting bot (500 contracts/day) and determine optimal architecture
- Detail governance specifics for OpenAI models on Bedrock, including regional deployment and regulatory implications
- Configure Bedrock Guardrails: content filtering, topic blocking, PII redaction, custom policies
- Explore GPT-OSS-Safeguard model for specialized content classification
- Know when to invoke Bedrock directly vs. using API Gateway
- Demo: Enable Model Invocation Logging and monitor Claude endpoint with CloudWatch
- Challenge Activity: Enable logging and monitoring on a Bedrock endpoint, review usage metrics, configure output guardrails
- Discuss integration patterns: Lambda, ECS/EKS, API Gateway, Step Functions
- Discuss how to make design trade-offs: model selection (GPT-OSS-20B vs. GPT-OSS-120B), region vs. latency, cost optimization, and resilience strategies (multi-region and failover)
- Understand the complete GenAI application lifecycle: data sources, model invocation, global endpoint design, monitoring, and feedback
- Demo: Walk through a complete gen AI application architecture diagram using all concepts covered in the course
- Capstone Activity: Design architecture for FinCorp Market Sentiment Analyzer with constraints: (A) Latency optimization, (B) Security (no public internet), (C) Accuracy (JSON format stability), (D) Safety (block insider trading topics/blacklisted stocks). Address: Network design, Model configuration (Temperature/Top P), Data strategy (RAG vs. Fine-Tuning), Reliability, Governance
