Gigi Voice Fix & Build-Safe Deployments — August 28, 2026

Today, I focused on enhancing the reliability of the gigi-voice system and addressing a critical audio capture issue that had been causing disruptions. My primary goal was to ensure that all build-time external calls were safe to prevent the daily deploy breakage that had been plaguing the site. After about an hour of concentrated effort, I managed to ship a significant improvement and a crucial fix that I believe will stabilize the user experience moving forward.

The first commit I made was a reliability improvement aimed at making all build-time external calls build-safe. This was an essential step because I was facing a recurring issue where the site broke almost daily due to the static build making live third-party calls during the prerendering of nearly 9,000 local leaf pages and sitemaps. Whenever any of those upstream services—like Overpass, Census, Wikipedia, open-meteo, or even Google—were slow or down, it would result in a failed build. This situation not only wasted precious time but also disrupted the deployment process, making it a frustrating experience for users and for me as a solo builder.

To tackle this, I implemented checks that ensure these external calls are reliable and won't break the build. It took some time to figure out the best approach, but I finally settled on a method that effectively isolates the build process from these external dependencies. I’m optimistic that this change will significantly reduce the frequency of deployment issues, allowing me to focus more on building features rather than constantly fixing broken builds.

The second commit was a fix for the gigi-voice audio capture issue, which had been the last remaining problem preventing the voice agent from functioning correctly. After resolving the Content Security Policy (CSP) token issue, I realized that the audio capture was still failing. The problem stemmed from the SDK attempting to load the rawAudioProcessor from an inline blob, which was getting blocked by the strict script-src settings. This led to the frustrating error message: 'Failed to load the rawAudioProcessor worklet module,' causing the agent to disconnect during sessions.

To fix this, I self-hosted the rawAudioProcessor AudioWorklet, similar to how I had previously self-hosted the audioConcatProcessor and libsamplerate. However, I had overlooked the input worklet, which was the missing piece. Once I identified this oversight, I implemented the fix, and I’m happy to report that the audio capture is now functioning as intended. This was a significant relief because it meant that the gigi voice agent could operate smoothly without interruptions, enhancing the overall user experience.

Working on these issues alone, with Claude Code as my AI assistant in VS Code, has been both a challenge and a rewarding experience. I often find myself reflecting on the journey of building this agentic web platform by myself. Each bug I face and each feature I ship brings me one step closer to my goal of creating a one-man-show company with a billion-dollar valuation. The lessons learned from these fixes are invaluable, and I’m excited to see how they contribute to the overall stability and functionality of the platform.

As I wrap up the day, I’m proud of the progress made. The fixes implemented today not only address immediate issues but also set the foundation for future enhancements. I look forward to continuing this journey, building out more features and improving the reliability of the gigi-voice system. The road ahead is long, but each commit is a step towards making the agentic web a reality. Tomorrow, I’ll dive deeper into the next set of features I want to implement, keeping the momentum going as I strive to build a powerful AI website builder that truly stands out in the market.