Conducting our own research about the laws surrounding marketing materials was critical to the future of the newly christened “Marketing Analyzer”. At the beginning of the project, we settled on using Large Language Models (LLMs) to replicate how HighCamp manually reviewed these documents. However, we were at a bit of a loss as to how to actually execute that strategy - talking with HighCamp about compliance often felt like we were trying to learn how to sprint when we had never walked before. Thanks to their patient guidance, some internal documents for their new employees, and a lot of time pouring over the SEC’s 400 page document on the “Marketing Rule” surrounding investment adviser marketing, we gained enough understanding about the problem to start creating the tool.

We had a lot of homework to catch up on.

In the time we’d been researching, we had also been able to settle on a final design plan for the Marketing Analyzer. HighCamp would upload a piece of marketing material (typically a PDF) to the Marketing Analyzer. After screening for as many compliance violations as possible, the tool would spit the document back out with comments added directly into the PDF - any requests necessary to make the marketing material SEC-compliant like so:

Thus, the critical question on our side was: “How should we start programming a way to screen for compliance violations?” After a little brainstorming, we began development with the ultimate goal of creating as much of a lift for the HighCamp team as soon as possible.

We broke down the process of evaluating marketing material into more and more granular pieces, and then individually analyzed each step to see what felt like a good fit for an AI. We ultimately settled on the three major components of the AI that would help HighCamp the most:

1. A knowledge of the more straightforward rules surrounding language in marketing materials. We wanted to ensure that an LLM could reliably find places where these rules were violated.

2. The ability to scan the document for disclaimers, to save them as reference for the rest of the document’s language.

3. Finally, the ability to review all of the text in the marketing material, checking for any violations with context from other parts of the document in mind.

A more in-depth breakdown of how we built the major 3 AI components

These manifested in the final backend architecture as 14 different LLM “chains” built to mimic a human’s line of reasoning. All of these chains are built to work together to reach our final goal: taking in text from marketing materials and writing down feedback if there’s reason to believe the text violates SEC compliance. Ultimately, all that was left for HighCamp to do was to:

1. Review that our comments were accurate, and revise any wording as desired.

2. Do a quick read-through of the marketing material to check if any information was misleading or blatantly untrue, based off of their deeper knowledge about their client.

An example from the final piece of our LLM pipeline, which generates feedback that will be commented into the document based on what parts of the “Marketing Rule” that the text violated.

Once we were confident with the backend’s performance, we bundled it up into an API in HighCamp’s Azure environment and started polishing up a user interface (UI) for the Marketing Analyzer!

The current Marketing Analyzer UI.

Just like the AI logic, a lot of thought went into the Marketing Analyzer’s UI. The HighCamp team needed to be able to see the original marketing material and create and edit comments in one platform. However, we didn’t want to stop there: we wanted to save the changes the HighCamp team makes to act as feedback which makes the AI logic behind the Marketing Analyzer better and better. Thus, we added some crucial functionalities to the UI to do so:

1. Clicking the Edit button allows HighCamp users to change the “compliance flag’s” classification, to better inform us how exactly the highlighted text violates SEC compliance.

   

   
   

   
   

2. In the same vein, clicking the Delete button asks HighCamp users to specify whether they felt the highlighted text was actually compliant with the SEC’s rules or if they’re deleting the comment for other reasons (relocating it, etc):
   
   

   

Finally, there were also some great quality of life additions that we were able to make for HighCamp’s ease of use:

1. Clicking on the “Jump to Text on PDF” button moves the marketing material to the relevant “compliance flag” highlight. The opposite also works - clicking on highlighted text in the PDF will shift the list of “compliance flags” to the relevant one!
   
   

   

2. Downloading the document saves all those compliance flags as comments on the PDF, and also quickly labels when the document was last reviewed by HighCamp.
   
   

   

All in all, the Marketing Analyzer is now at the point where the whole team at HighCamp can get some great use out of it while we continue making more custom products for them! We’re looking forward to showing off some more of those tools and how the Marketing Analyzer evolves in the future!