Local Fetch Enhancements & Tavus Fixes for Spatial AI — September 4, 2026

Today, I focused on refining the local fetch capabilities and fixing some quirks with Tavus, all while enhancing the overall user experience on my agentic web platform. I shipped one new feature and resolved two critical issues in about an hour of concentrated work. The improvements matter because they directly impact how users interact with the AI tools and the overall efficiency of the platform.

The standout feature I introduced today was the untrackedFetch function for external render-path calls. This enhancement is crucial for optimizing how our platform interacts with various data sources without compromising performance. It’s a step towards making the AI website builder more responsive and efficient when fetching external data. The goal was to streamline this process, ensuring that the user experience remains smooth and that the platform can handle requests without unnecessary delays.

However, the journey to get this feature live wasn’t entirely smooth sailing. I spent a good chunk of time diagnosing some issues that arose during initial testing. It turned out that the untrackedFetch wasn't handling certain edge cases well, particularly in scenarios where external API responses were slower than expected. I implemented a few quick fixes that allowed the function to retry fetching data and gracefully handle timeouts. This not only improved the reliability of the fetch process but also taught me a lot about error handling in asynchronous calls.

On the fix side, I tackled a couple of pressing issues that were affecting the Tavus integration. The first fix involved unblocking in-app frames for the AISEO board and Calendly booking. This was a critical issue, as users were unable to access these features seamlessly within the app. After reviewing the interaction configurations, I found that certain nested frames weren’t rendering correctly due to some misconfigurations. It took some digging, but I managed to get everything back on track, ensuring that users could access the full functionality of these tools without any hiccups.

The second fix was related to the Lucas radio feature. I discovered that the Lucas live session was mistakenly using the stock id for the Interviewer persona instead of the owned one, which has 28 tools integrated. This was a significant oversight as it limited the capabilities available to users during live sessions. I quickly rectified this by updating the configuration to ensure that the correct persona was utilized. Not only did this fix improve the overall functionality of the radio feature, but it also aligned with my goal of creating a more agentic web experience where users can leverage the full potential of the tools at their disposal.

Reflecting on today’s work, I’m proud of the progress made but also mindful of the challenges that arise when building a platform like this solo. It’s a constant balancing act between adding new features and ensuring that existing functionalities operate flawlessly. Using Claude Code in VS Code has been invaluable in this process, allowing me to iterate quickly and leverage AI assistance without incurring the costs of collaborative tools. I can’t help but think about how each of these improvements brings me closer to my goal of building a one-man-show company with a billion-dollar valuation.

As I wrap up for the day, I feel a sense of accomplishment in shipping meaningful updates that enhance the user experience on my platform. Every fix and feature is a step toward creating a more robust and agentic web environment. I look forward to tackling the next challenges that lie ahead, knowing that each day brings me closer to realizing my vision.