Over the last four months, my focus and projects have evolved rapidly as I have navigated different tools, ecosystems, and real-world friction points.
April: The Education Phase
I spent my first month deep in learning and discovery. I spent a good deal of time on Coursera looking for courses that would increase my AI knowledge and sharpen my coding skills. During that time, I managed to earn my Google AI Professional and Google AI Generative Leader certifications, along with a number of other smaller coding certificates.
But the real practical learning came from toggling directly between the Google and Anthropic ecosystems to understand how their products work under the hood. I spent time evaluating Google’s Gemini, NotebookLM, Google Cloud, Gems, and Antigravity, alongside Anthropic’s Code and Cowork. I also set up Pro accounts for both platforms to understand the practical economics of their billing, pricing, and programmatic API usage.
May: Prototyping and Execution
May was about taking that baseline knowledge and actually building things. I relied heavily on Google’s Antigravity IDE to ship my first AI projects. These ranged from small utilities — like a Google Docs extension and a workout timer — to a much more ambitious attempt at building an agentic development team and my own agentic OS.
To test the boundaries of these tools, I built two versions of the OS: one leveraging Google Antigravity, Google’s Agent Development Kit, and Google Stitch for prototyping, and a second version using Claude Code and following a self-paced build program from AIDB Training to compare their capabilities. It was also at this time that I also started prototyping a consumer app, which I plan to detail in a future post.
June: Real-World Automation ROI
In June, I put my pure AI projects on hold to take on a volunteer coordination role for a non-profit organization. The system they had in place was highly bespoke, maintained by a single individual, and practically impossible for an outside audience to comprehend.
I decided to rebuild the entire architecture using Microsoft 365 and Power Automate. The goal was to remove human error and drastically improve efficiency. After a solid month of work, the new automated system was in place.
By applying the right automation, I reduced a daily four-to-five-hour administrative slog down to just one hour of required maintenance.
July: Building Wren
By early July, I returned to my core AI roadmap. The majority of my current development is focused on building out my own personal AI agent named Wren, alongside an LLM wiki powered by Obsidian.
Wren initially started out as a simple way to automate my morning brief and act as a basic organizing assistant. But as I have continued to iterate, she is rapidly developing into a far more capable personal agent, complete with her own memory and learning system. Both the wiki and the agent are very much works in progress, but the daily capability gains are impressive.
This is where my focus remains today — iterating on Wren, refining local models, and pushing the boundaries of what these systems can actually execute on my own hardware. I will share a much deeper look into Wren’s development and specific capabilities in an upcoming post.