Connecting thoughts

We found two keys to answer this question while building a full-featured coding agent in just 250 lines of Ruby code.

While working on RubyLLM::Monitoring, we needed a way to instrument all RubyLLM operations. But we wanted to do it without changing RubyLLM. Read along to know how we did it.

When you’re using multiple LLM providers, tracking costs manually becomes impossible fast. We needed visibility into our AI spending and LLM performance. Here’s the monitoring engine we built for Rails.

Learn how to integrate Model Context Protocol (MCP) with Rails to create AI-powered conversational interfaces that transform traditional web applications into intelligent, chat-based tools.

We replaced a broken similarity search with AI that understands context and intent over specs. Results went from chaos to relevant overnight.

Learn how we improved image quality for thousands of boat listings using artificial intelligence, with the objective of improving the user experience and conversion rates.

Learn how we used multi-modal Large Language Models to automatically categorize more than 1 million boat images, reducing months of manual work to a couple of days.

The most important part of making OpenAI’s batch processing API work in the real world is building a reliable polling system. This post explains why that’s necessary, and what else you’ll need to handle: token limits, partial failures, and retries.