Every VC I talk to lately is building something. A partner shows me a dashboard he put together over a weekend with Claude or Cursor: his portfolio, his marks, his own view of the book. I love it. Founders live in these tools now, and an investor who can say "I build with the same stuff you do" gets a different kind of conversation. Portfolio monitoring is the obvious place to start. It's your data and it's a contained problem.
We love this use case, and it's exactly what we built GoodStream to feed. So build it. The workflows are the fun part, and they should be yours.
The hard part is the data those workflows run on. Every number in that dashboard came out of a document somewhere, and somebody has to pull it out of the paper. The price per share came off a stock purchase agreement, the liquidation preference off a charter, and your ownership percentage off a cap table that was correct on the day it was sent and has been drifting ever since. Today that somebody is a language model. In our testing, a single extraction pass gets the number right about 63% of the time, even with a prompt tuned for that exact field. The other third of the time it's confidently wrong, which in practice means it returns a price per share to four decimal places and the number is from the wrong series, or the wrong document, or it doesn't exist anywhere in the paper at all. In a demo you spot check a handful of numbers, they look right, and you move on, and the LP who gets the number later doesn't spot check it.
That's fine for a demo. It isn't fine for the numbers you put in front of your LPs.
The truth in private capital is in the documents, not in the ledger. Getting that truth out takes two things, and neither one is a weekend project.
Accuracy
The first is accuracy. We run a separate extraction agent for every field, more than one model reads each one, and when they disagree a person settles it with the source page open. Every edge case we hit gets folded back into that agent, so the next fund never sees it. After a few thousand documents, the edge cases are where all the value sits. Being able to properly understand the side letter that quietly changes the preference isn't a nice to have, it's a requirement, and it takes a lot of side letters before an agent reads that one correctly the first time.
Data association
The second is data association, and this is the one people miss. Knowing you hold a convertible note from 2025 is easy. Knowing it converted into the Series A in 2026, what it converted into, how much you actually own now, that the original cost isn't counted twice, and that the accruing dividend is accrued properly, that's a chain of documents talking to each other. The note is one document, the conversion notice is another, the Series A SPA and the amended charter are two more, and the cap table that comes out the other side is a fifth, and a dashboard that reads each of those on its own will show you the note and the Series A shares side by side as if you own both. Each of those documents is correct on its own, and the error only shows up when you line them up against each other. No home-grown, LLM based extraction gives you that, because it reads one document at a time and association is reconciliation work across all five.
So build the dashboard. Build the LP portal, the pacing model, the thing your partners keep asking for. Just point it at data that's been rigorously processed, argued over by a human, and traced back to the paragraph it came from. That's the part we do all day, so the fun part can stay yours.
