This post explains how to use synthetic data to bootstrap evaluations for Retrieval-Augmented Generation (RAG) systems, including setup, data generation, and performance optimization techniques.
About:
Dylan Castillo is an independent AI consultant with a decade of experience, passionate about AI, machine learning, side projects, and technical writing.
Website:
Specializations:
Interests:
Outgoing Links:
Subscribe to RSS:
Gemini 1.5 Flash's structured outputs show mixed performance, with significant issues in constrained decoding, highlighting the need for personal evaluations in choosing output methods.
Explore how to implement parallelization and orchestrator-worker workflows using Pydantic AI to enhance task efficiency and management.
This tutorial outlines how to build ReAct agents using LangGraph, highlighting their autonomous decision-making capabilities and contrasting them with agentic workflows.
The post contrasts agentic workflows with agents, detailing their appropriate applications and introducing LangGraph for building structured workflows with LLMs.
Temperature and seed parameters in LLMs control output randomness and reproducibility, affecting how models generate text based on token probabilities.
RAG improves LLM responses by integrating external information through a structured pipeline, enhancing accuracy and efficiency in document retrieval and response generation.
The post outlines the implementation of prompt chaining with Pydantic AI to enhance content generation workflows through a structured approach.
Building a ReAct agent with Pydantic AI involves setting up asyncio, integrating Logfire for monitoring, and implementing tools for weather retrieval and guideline checking.
A tutorial on using LangSmith to effectively monitor and evaluate LLM applications, highlighting essential setup and evaluation techniques.
This post explores building an evaluator-optimizer workflow with Pydantic AI, demonstrating its application in generating and refining content on hard drug consumption.
A detailed guide on implementing a routing workflow with Pydantic AI to classify and manage user queries through specialized agents.
Mastering function calling and structured outputs enhances LLM applications, enabling them to interact with external tools and produce structured data effectively.
Key parameters for building LLM-based products are discussed, including model selection, temperature control, and response formatting, with practical usage suggestions.
Effective prompt writing for LLMs hinges on clarity, structure, and specific instructions, as outlined in six key principles.
The post reveals that Japanese is the most expensive language in terms of input tokens, requiring significantly more tokens than English for the same text.
Dylan Castillo's annual review highlights his journey through personal and professional challenges in 2024, emphasizing lessons on focus, intensity, and the importance of staying active.
CYA is a crucial communication strategy in the workplace that helps individuals clarify priorities and mitigate blame in project failures.
4.5ClaudeCodeholic – Dylan Castillo
2025-10-18 • productivity mental health code ai
Dylan Castillo shares his addiction to Claude Code, emphasizing its comfort and emotional relief over traditional productivity measures in coding.
Dylan Castillo's annual review reveals a successful year in consulting, personal growth through challenges, and ongoing struggles with stress and work-life balance.
Dylan Castillo expresses frustration with the saturation of low-quality AI projects while acknowledging the positive impact of AI on his work and creativity.
...