They say no plan survives the battlefield, but hey, we're not going to let that stop us now. First things first, what are we building? It's what everyone else is building: a research agent, specifically for the fintech market. If you have worked in fintech, you're aware of customer due diligence and just how much work goes into it. We're setting out to try and simplify that work.
Yeah, I know... I know it's probably not gonna work—it's a high-accuracy industry where the details matter, and judgment goes way beyond just sentiment analysis of search results. Rather, we need to find different data points and judge based on how these data points relate to each other. Just because something has a negative sentiment in one context doesn't mean it's always bad—sometimes it's just okay. Other times, even positive sentiment data can be judged as negative depending on factors like timing and relationship with other data.
How are we doing this?
You might have noticed we haven't discussed anything about frameworks, which LLM providers, or versions we're going to use. At this point, I think most LLMs are mostly on par with each other—at least for our current use case. What matters more actually is how well you understand the problem space you're attempting to solve. The more context you have on the topic, the better. Having subject matter experts is even better—at least that's my assumption going in. We'll see how much that changes.
So what do we actually need? We want to be able to provide a company or person name, and the "agent" should be able to identify the company, any number of people associated with it (at least the top-level C-executives), see if they're connected with other companies significantly enough to warrant looking into, financial trends of said company, and other relevant data points—all combined to give a score. So we essentially need three things: a data source, some way to judge the data, and a way to find connections between data points.
Now we could complicate things by adding more data points, but hey, it's our first challenge—let's take it slow. With that in mind, our first version is going to be quite simple, just to try and get a feel for what's possible. We'll have a data source, search, and an LLM judge.
flowchart TD
A[Input: Company/Person Name] --> B[Data Collection]
B --> C[Company Lookup]
B --> D[Key People Lookup]
B --> E[Basic Financial Data]
C --> C1[Company Profile & Status]
D --> D1[C-Level Executives]
E --> E1[Revenue/Financial Health]
C1 --> F[Simple Scoring Engine]
D1 --> F
E1 --> F
F --> F1[Company Health Score 1-10]
F --> F2[Leadership Score 1-10]
F --> F3[Basic Risk Level]
F1 --> G[MVP Output]
F2 --> G
F3 --> G
G --> H[Simple Report:<br/>- Overall Score<br/>- Key People Found<br/>- Financial Status<br/>- Risk Level]
style A fill:#e1f5fe
style B fill:#f3e5f5
style F fill:#e8f5e8
style G fill:#fce4ec
style H fill:#e3f2fd
What about the tools?
I guess you couldn't wait to get to this part. There's nothing grand here, really..I'm leaning towards the OpenAI JS SDK and the Perplexity API. See you in the next one!