Amazon Bedrock: foundation models as a service

Lesson 3 of 4 in The ML Lifecycle on AWS.

Amazon Bedrock is the other engine, built for the half of the exam that did not exist in MLA-C01: foundation models. The pitch is one serverless API over many hosted FMs — Amazon's own Nova family plus models from Anthropic (Claude), Meta (Llama), Mistral, Cohere, and others — with no instances to size, scale, or patch. You choose a model the way you choose an instance type: by capability, context needs, latency, and token cost.

Around the models, Bedrock layers the production machinery MLA-C02 tests directly: customization (fine-tuning and Custom Model Import), grounding (Knowledge Bases for RAG), safety (Guardrails), measurement (model and RAG evaluations, including LLM-as-a-judge), and agents — with Bedrock AgentCore supplying the managed runtime and observability to operate them.

Model choice

One API, many FMs: Amazon Nova, Anthropic Claude, Meta Llama, Mistral, Cohere, and more. Pick by capability, context window, latency profile, and cost per token — and swap models later without re-architecting the app.

Customization

In rising order of effort: prompt engineering; fine-tuning on labeled examples; continued pre-training on raw domain text; Custom Model Import to serve open-weights models you already trained through the same Bedrock API; and distillation to train a smaller, cheaper student from a larger teacher.

Knowledge Bases (RAG)

Point it at your documents (S3 and other sources); it chunks and embeds them into a vector store it manages (Amazon OpenSearch Serverless by default — or one you bring) and retrieves the best chunks at query time. The managed path for "make the model answer from our documents."

Guardrails

Configurable policies: content filters, denied topics, word filters, sensitive-information (PII) handling, and contextual-grounding checks. They apply to inputs and outputs, and the ApplyGuardrail API extends the same policy to models running outside Bedrock.

Evaluations

Managed evaluation jobs for models and for RAG pipelines: automatic metrics, human-review workflows, and LLM-as-a-judge scoring. When a Domain 2 scenario asks how good is this model?, this is the managed answer.

Agents + AgentCore

Bedrock agents let an FM plan multi-step tasks and call tools and Knowledge Bases. AgentCore is the production layer: a serverless, framework-agnostic runtime for agents (whatever framework they were built with), plus session memory, identity, a tool gateway, and observability dashboards that trace every step. It's on the C02 in-scope list — know what it is.

Key terms: Amazon Bedrock, Foundation model (FM), Retrieval-augmented generation (RAG), Bedrock Knowledge Bases, Bedrock Guardrails, Bedrock AgentCore, Custom Model Import

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.