Elizabeth Fuentes Leone is an AI engineer and developer advocate focused on building reliable AI agents and translating machine-learning concepts into practical, reproducible software. At AI Engineer World’s Fair 2026, she represented Amazon Web Services as a developer advocate working on agentic applications.
Fuentes Leone studied data science at Universidad del Desarrollo in Chile and began publishing Spanish-language cloud and machine-learning tutorials before joining AWS’s Latin American developer advocacy team. Her early work included a hands-on introduction to Amazon SageMaker and a serverless system for loading CSV data into DynamoDB. At AWS, she developed a personalized anime-recommendation project that extended beyond model training to APIs, web applications, authentication and engagement metrics.
Her subsequent Strands Agents examples explore multimodal applications, tool integration and deployment. A central conviction shapes her agent engineering: reliability depends on architecture and executable controls, not increasingly elaborate prompts.
How she makes agents dependable
- Semantic tool selection: Embeddings and top-k retrieval expose only relevant tools for each request, reducing unnecessary context and limiting opportunities for incorrect tool calls. Conversation history persists while available tools change between invocations.
- GraphRAG for verifiable answers: Neo4j knowledge graphs and Cypher queries support counts, aggregations and multi-hop reasoning across complete datasets where retrieving a few text passages produces incomplete or speculative answers.
- Executor-validator-critic workflows: Separate agents perform an operation, check its outcome and approve or reject the result, preventing failed bookings from becoming fabricated confirmations.
- Code-enforced business rules: Application hooks enforce capacity limits, reservation requirements and verified payment before tools execute; consequential constraints belong in code, not prompt text.
- Runtime steering: Agent Control SDK redirects recoverable policy violations toward valid alternatives, such as splitting a large group across rooms, while preserving hard stops for requirements such as payment verification.
Her AI Engineer World’s Fair session on agent hallucinations connects these techniques to Strands Agents, OpenAI, Amazon Bedrock and Ollama, while distinguishing simplified local demonstrations from production systems requiring managed routing, memory, observability and enforceable policies.
