Hermae

Challenge

Design-system maintainers at large companies kept asking how to get a new team onto the system without a long training cycle. The source of truth was already there. What was missing was an interface that could retrieve against it and answer in the language of the team asking.

Public demos in 2024 drew interest from Fortune 500 maintainers, which isn't a sale. The product still had to sit inside security review, connect to a customer's own system, and stay useful on a Tuesday when nobody from our side was on the call.

Approach

We blended LLM generation with living design systems. Hermae retrieved against system source and generated an assistant on top, so answers came from the customer's tokens and docs instead of a generic model of what a design system is.

We shipped an API and a dashboard so teams could connect their own assistants. When the product needed more hands we scaled the build to about ten contractors. The technical work held: RAG and vector-store patterns we still use, and a custom LLM app you could point at a real corpus.

The commercial side didn't get the same discipline. Every serious conversation entered enterprise security and compliance review, and we kept building for the next demo instead of owning the buying conversation. We'd already seen that failure on Arcade.

Outcome

The product held up and the sale didn't. For most buyers the tool stayed useful but optional once procurement got involved, and optional tools die in a security queue. We wrapped the company in February 2025.

What's still useful is the RAG and vector-store work, plus a clearer read on when a well-built AI product still can't clear an enterprise buying motion. That lesson sits behind later agent systems work: retrieve against a corpus the customer already trusts, and don't confuse a working assistant with a product someone is allowed to buy.