Santiago Paz, homeCV download (PDF, 159 KB)
All work

bedrock-genai-labs

A study curriculum I assembled for the AWS generative AI certification: 26 lab scaffolds with reference solutions, covering retrieval, an agent loop, guardrails and LLM-as-judge evaluation. I am still working through it.

  1. D1

    Foundation models and retrieval: embeddings, chunking, building a FAISS store by hand, hybrid retrieval with reranking.

  2. D2

    Implementation: streaming, resilience, tool use, writing an agent loop by hand, model routing.

  3. D3

    Safety and security: guardrails, PII redaction, prompt injection, grounding checks.

  4. D4

    Optimization: token accounting, cost, prompt caching, monitoring.

  5. D5

    Testing and evaluation: LLM-as-judge, retrieval evaluation, prompt regression.

26 lab scaffolds and 64 quiz questions, organized by the five exam domains. The exercises are still unfilled; I am working through them.

Layers built

  1. Shared tooling
  2. Labs
  3. Solutions
  4. Progress tracker

Role

Sole author of the curriculum: the lab scaffolds, reference solutions, shared tooling and a progress tracker.

Stack

Python, boto3, Amazon Bedrock, FAISS, BM25, Strands Agents SDK and Bedrock Guardrails.

Study material in progress, not shipped work. I wrote the 26 scaffolds and their reference solutions, and the exercises still carry their unfilled TODO markers. This is the syllabus I am working through for the certification. It is not production RAG, agent or Bedrock experience, and I do not present it as any.

Highlights

  • Written to build the primitive before reaching for the abstraction: a vector store before FAISS-as-a-service, an agent loop before an agent SDK. Knowing what a framework does for you is what makes it debuggable later.
  • Retrieval from the ground up across the labs: chunking strategies, embedding generation, a FAISS vector store written by hand, then hybrid dense and BM25 retrieval with reranking.
  • Safety and evaluation get their own domains: Bedrock Guardrails, PII detection and redaction, a prompt-injection suite, then LLM-as-judge scoring, retrieval metrics and prompt regression testing.
  • Shared infrastructure for the Bedrock client, config, token pricing and a synthetic corpus to retrieve against, so each lab starts at the exercise rather than at the setup.

Overview

A study curriculum I assembled while preparing for the AWS Certified Generative AI Developer - Professional (AIP-C01) exam.

26 lab scaffolds organised by exam domain, each a set of TODO(you) exercises with a reference solution written and held back. The five domains cover foundation models and retrieval, implementation (streaming, resilience, tool use, agents, routing), safety and security, optimization, and testing and evaluation. Written in Python and boto3, with shared infrastructure for the Bedrock client, config, token pricing and a synthetic corpus to retrieve against.

The recurring choice is to build the primitive before reaching for the abstraction: a vector store before FAISS-as-a-service, an agent loop before an agent SDK. This is the syllabus, and I am still working through it.