PAND·AI Gets a Brain Upgrade

When we launched PAND·AI in December last year, we mentioned that it was bound to evolve in its first year. Part of that was about gathering your feedback, which was indeed helpful, especially in the first few months after launch when we fixed some quirks in the setup. The other part was due to seeing how fast AI models were improving at the time. Almost a year later, we can confirm that they improved even faster than anyone expected.

The models today can reason more carefully, hold much more context while staying consistent, and are much less prone to the once-dreaded hallucinations we warned about in our launch post. Now felt like the right time to put those improvements to work, and the new version is currently live. In fact, some of you may have noticed that since last weekend, opening a report regenerates it in full rather than simply updating the Technicals and Valuation sections. This is simply because PAND·AI, as of last weekend, is using its new brain to update the reports to the newest version.

If you load a ticker that hasn't been updated yet, expect a few minutes' wait the first time. If the report is already available, it means its newest version has already been generated.

What hasn't changed

Even with a new brain, PAND·AI remains what it was at its core. We do not simply ask a model to generate a report from data sourced no one knows where. We provide the model with data validated from our trusted providers, pre-compute any necessary metrics, and provide guidelines through multiple back-and-forths with AI to finally come up with the final report. We use financial reports, earnings call transcripts and multiple statistics, some of them our own, to guide the report towards what is most relevant.

This structured process also helps us increase the quality by designing the report around how we analyze a company as investors. Each section is crafted to answer a specific question, from the obvious "what do the financials have to say" to the more subtle "what is management trying to put in place". It also allows us to check each output and re-run a section if it falls short before it gets shown.

What you'll notice

The biggest change, in my opinion, is how much more cohesive the report is as a whole. This is subtle and subjective, of course, so we're interested in hearing what you think, but the reports now read much more as one piece than a bunch of parts glued together. The SWOT and bull/bear case feel more like a conclusion to the rest of the report than added independent sections.

Speaking of the SWOT, we've made it sharper. Each point is now a headline accompanied by a supporting explanation, and each quadrant has its points ranked from most to least important. The new models seem to have a better understanding of how to rank them. Additionally, most SWOTs now fit on a single page, with the occasional one splitting into two. In the previous version, some SWOTs even took three pages, which did not make for a nice reading experience.

There's also a bunch of smaller things. For example, we made sure the model remains consistent between sections. Sometimes, the transcript mentioned a value that was not exactly the one from the financial reports. The model now knows to favour the one from the financials on those small discrepancies.

We also improved how young or transformed companies are assessed in the Valuation in Context section. This section looks at a stock's historical valuation to judge how cheap or expensive it currently is, as opposed to looking at the valuation numbers in a vacuum, so a short history can make the comparison irrelevant. This is especially true for companies that were not profitable at first, or just barely, since their valuation multiples were either really big or simply not valid, making the comparison moot.

What's next

This is our last round of updates on PAND·AI before we drop the Beta label. We mentioned in the launch post that we would drop it once we were confident in PAND·AI's robustness, and this version gets us there. We welcome any feedback you may have on it so we can make the final tweaks. Please reach out at dev@thewealthumbrella.com with any comments or suggestions.