In 2023, Maya spent the better part of a year as the head of investor relations for a mid-market private credit fund. Her job, on paper, was to communicate with limited partners. Her job, in practice, was to be a human firewall — re-reading every outbound message, rewriting half of them, and apologizing for the rest.
The thing nobody told her in the offer letter was how much of investor relations is regulatory rather than relational. Performance figures had to be carefully framed. Forward-looking statements needed boilerplate disclaimers. Anything that touched accredited-investor status required verification. The PPM was a 134-page document she was supposed to know by heart and quote from never.
And then, in late 2023, an analyst on the team pasted a passage from the PPM into a ChatGPT window to draft a response. The compliance officer saw the audit log the next morning. The fund settled with the regulator quietly. Nobody got fired but everyone learned a lesson, which was: the model is not the problem. The lack of constraint around the model is.
The naive thing didn't work.
Maya's first instinct, like everyone's, was to write a long policy document. Train the IR team harder. Add a checklist. The checklist didn't survive contact with a slow day and a chatty LP.
Her second instinct, like everyone's, was to ban the chatbot. That worked for about three weeks. Then someone needed to draft a follow-up at 11pm on a Thursday and the rules quietly relaxed.
“The answer wasn't a better prompt. It was a different shape of system — one where the model couldn't do the unsafe thing even if it wanted to. Where the constraints lived in the database, not in the team's discipline.”Maya Okonkwo, co-founder & CEO
She emailed Daniel — a friend from law school who had spent five years at the SEC's Division of Enforcement. They spent six weekends sketching a different design: a chat surface where classification ran before retrieval, where retrieval was filtered at the SQL layer rather than in application code, where the audit log was an append-only table with a database trigger refusing to mutate it. The model was just the last step.
What we built first.
The first version of Dalphe was a Python script and a Postgres database. It classified messages with Claude Haiku. It refused anything it wasn't sure about. It wrote everything down. Maya's old fund was the first customer. The compliance officer said, in writing, that he'd never seen an AI tool he was willing to defend to an examiner. He was willing to defend this one.
We've kept the shape since then. The pipeline still runs in the same order — classify, gate, retrieve, compose, audit — because that order is the product. The audit log is still the part our customers love most, which we did not predict.
Where we are now.
Twenty-three private fund managers run their IR channel on Dalphe today, representing about $14 billion in combined AUM. None have been fined. Several have been examined. All passed with the audit log as their first exhibit.
We're hiring for engineers who care about correctness more than velocity, and a founding solutions person who has sat across the table from a SEC examiner and can explain what makes them comfortable. If that's you, the email at the bottom of this page is real.