Building an Alternative to Examine.com: A Challenging Journey!

Many of you might already know about Examine.com, a leading resource for supplement research. Examine focuses on specific health areas, scouts for related research papers, and then describes the link to different supplements. This meticulous process is understandably labor-intensive, which is why they charge a fee for accessing their data. However, with the advent of AI, many parts of this process can be automated, and that's the solution I'm working on.

For instance, here's a compilation of research linked to Reduced Body Weight:

https://pillser.com/health-outcomes/reduced-body-weight-158

I've indexed thousands of research papers and extracted insights that connect these studies to various supplements on the market. My long-term goal is to create a pioneering supplement store where every product is linked to research. Unlike Examine, I plan to make all research summaries public and instead focus on earning affiliate revenue from related supplement sales.

The biggest challenge is ensuring data accuracy. Given the complexity of these topics, I am currently limiting insights to demonstrate a link between the study, health outcome, and substance. Users are then directed to the actual research papers to build confidence in their decisions. However, as AI models evolve, I aim to expand this into a comprehensive insight engine.

AI and large language models (LLMs) are crucial in this process. Finding research papers is relatively straightforward with resources like PubMed. I use a combination of API services, varying in cost and speed, to scout for relevant mentions in the research (using fast and cheap models), validate the relevance of these mentions (using top models), and finally ensure the accuracy of the summary versus the excerpt (using another model). The idea is that the first model may make mistakes, the second filters out false positives, and the third acts as a final safeguard.

I am deeply fascinated by this problem domain and, more broadly, by data normalization and the applications of LLMs in solving these problems. While this project is currently a hobby that I anticipate will be a money-losing activity for a long time, I believe there is a significant chance that Pillser could become a preferred site for supplement buyers due to its unique combination of scientific backing and extensive inventory.

I am probably a few months away from completing the database, but I wanted to share this for early feedback. The website is called Pillser, and you can already search for different health goals and associated research directly from the landing page.