Humanizer Loop Enhancements & Local POI Database Updates — September 28, 2026

Today, I set out to improve the humanizer loop and the local POI database, essential components in my quest to build the agentic web at hashtag.org. After a long 9-hour day, I shipped 26 commits, including 17 features and improvements, and 9 fixes. The work involved refining how the humanizer processes content and enhancing our local business listings by lifting caps on the POI database.

One of the standout features I implemented was the ability to tag each job processed by the humanizer with the corresponding page URL. This enhancement allows me to track which URLs are being humanized, providing valuable insights into the performance of the content generation process. However, I encountered some challenges along the way, particularly with the humanizer's ability to handle number-heavy paragraphs, which often led to fabrications. I had to revert to a strict single-pass rewrite approach to maintain quality and avoid generating AI-like content.

The local POI database updates were a significant lift today. I raised the cap on business listings from 200 to 5000, allowing us to unlock the full potential of our 4 million-row POI database. This change was crucial for serving dense metropolitan areas accurately. However, implementing this change wasn’t straightforward. I had to ensure that the sitemap reflects all city-industry combinations and not just those that were drip-activated, which required a complete overhaul of how listings were generated. It was a real grind, but seeing the potential for improved local search visibility made it worth the effort.

Another key fix involved the sitemap generation for hashtags. I realized that we were listing tags based on their presence in enrichment files rather than their actual usage. This led to a bloated inventory that didn’t accurately reflect what users were searching for. By focusing on real usage, I was able to cut down the list significantly and provide a more authentic representation of our offerings.

Performance improvements were also a big part of my day. I optimized the humanizer loop to stop early when a high confidence score is detected, which cut down latency without sacrificing quality. It was a delicate balance to strike, but the results showed a clear reduction in processing time. Additionally, I made significant strides in ensuring that the humanizer only generates content on our 4090 node, eliminating reliance on local transformations that were leading to inconsistencies.

As I worked through these improvements, I couldn't help but reflect on the solo nature of this journey. With Claude Code as my AI assistant in VS Code, I’m able to tackle these complex challenges without the overhead of a larger team. This solo approach not only keeps costs down but also allows me to maintain a clear vision for the project. Each commit is a step toward my ambition of creating a one-man show with a billion-dollar valuation.

Despite the progress, I did encounter a few bugs that ate into my time. For instance, I had to deal with a significant issue where the humanizer occasionally generated fabricated addresses and phone numbers. This problem stemmed from the way facts were being checked against the original inputs. I implemented stricter checks to ensure that only accurate information is retained, but it required a lot of trial and error to get right.

Overall, today was a productive day filled with meaningful improvements and valuable lessons. Each feature and fix not only enhances the product but also brings me closer to realizing the vision behind hashtag.org. As I continue to build out the agentic web, I’m excited to see how these enhancements will improve user experience and drive engagement. Tomorrow, I plan to dive deeper into optimizing our AI content detection capabilities, ensuring we're at the forefront of innovation in this space. With each passing day, I feel more confident in the direction I'm headed and the impact this platform can have on the future of online interactions.