I want to share my experience in the cloud, AI, and agents development areas.
I built a local-first Python CLI that reviews unpublished blog drafts with Claude and Codex in sequence. The tool keeps orchestration, reports, and publishing manual, while supporting mock mode for no-network testing. This post explains the workflow, privacy boundaries, provider authentication, and why human review remains essential.
In this post, I explore a lightweight handoff pattern for coordinating GitHub Copilot, OpenAI Codex, Claude Code, and IBM Bob in one local-first Python project. The goal: reduce context loss, make agent work traceable, and turn AI-assisted coding into a more structured engineering workflow.
This post compares various AI coding assistants/agents, emphasizing privacy from a developer's perspective. It highlights how modern AI systems function as integral tools in software development, shifting the privacy discourse from mere data training concerns to broader issues like data sovereignty, operational exposure, and pricing implications, particularly for individual developers.
This post reflects on Contextual Retrieval with Milvus in RAG systems. It explains how generated context can improve chunk retrieval, but also changes the retrieval corpus. Once generated context is indexed, validation, traceability, and quality control become architectural responsibilities—not optional implementation details.
The article continues the exploration of the AI Operational Complexity Cube, emphasizing that modern AI systems encompass software, infrastructure, and probabilistic AI components. It highlights the need for comprehensive testing approaches that consider interactions across these dimensions, as proper evaluation requires observing behaviors that emerge from integrated systems rather than isolated code.
AI systems rely heavily on data quality, which is often overlooked despite modern technical architectures. Issues like outdated, incomplete, or misaligned data can undermine system reliability, regardless of the sophistication of the components. Effective AI requires both high-quality data and solid technical infrastructure to meet user expectations and ensure trust.
AI is transforming the software development lifecycle, shifting focus from coding to reviewing AI-generated systems. While AI tools simplify software generation, building trustworthy systems remains complex. Traditional review processes may no longer suffice. This raises a critical question: how can humans responsibly.
This post discusses a local setup utilizing IBM Bob to generate an agent for watsonx Orchestrate, specifically with tools from the Galaxium Travels MCP server. It explains the architecture, customization of Bob, and integration with various components, providing both learning and practical implementation value for developers.
This blog post explores the paradox of AI's growth potential versus the increasing trend toward data protectionism. It highlights how AI tools are hindered by data access limitations, posing risks to innovation. The observation implies that as data becomes more valuable, organizations may withhold it, undermining the openness that has historically fueled AI development.
The article explores the potential for using the Model Context Protocol (MCP) as a primary backend interface instead of traditional REST APIs in AI-enabled applications. Through the Galaxium Travels experiment, it examines the advantages and disadvantages of an MCP-first architecture, advocating for its use to reduce duplication and complexity while acknowledging REST's established role in many ecosystems.
Thomas Suedbroecker's Blog 