How a tool we built for ourselves became an early warning system for funds

Olma started as an internal tool. We dictate a lot, we were paying for Wispr Flow to do it, and the thing we actually wanted was narrower than what we were renting: press a key, talk, have the text land in whatever field has focus, and keep the audio on the machine.

So we built it. Olma is a macOS menu-bar app for Apple Silicon. Transcription runs on device through Parakeet, rewriting runs on device through Apple's Foundation Models, and nothing is sent anywhere unless the user turns a cloud feature on. It is signed, notarised and sold directly.

The interesting part was not the dictation. It was what we had to build underneath it.

Modes turned out to be the product

To make dictation useful we gave it modes. A mode is a small definition: a name, a prompt, worked examples of input and output, a language, and where the result should go. Speaking into an email mode produces something different from the same words spoken into a notes mode.

Once modes existed, the app stopped being a dictation tool and became a way to turn speech into a structured artefact. The transcript was never the point. The shape you put it in was.

That reframing is what made the second product possible.

The problem funds actually have

Funds find out late. A portfolio company is struggling in March and the fund hears about it in the June board pack, by which point the useful window has closed. Not because founders are hiding things, but because writing an honest update is work, and work that reports a problem is work nobody wants to do on a Friday afternoon.

The information a fund needs most is exactly the information that is hardest to ask for: what is blocked, what is being avoided, what the founder has not solved yet. Structured reporting collects the opposite. It collects what looks defensible in a form.

Journalling as the interface

So we changed the ask. Rather than a form to complete, an executive team gets a journalling prompt at the end of the day, week, month or quarter, and they talk for two minutes.

Talking is the mechanism, not a gimmick. People report problems out loud that they would never type, because typing invites editing and editing invites tidying. A spoken end-of-week reflection tends to contain the real thing: the hire that is not working, the customer who went quiet, the decision that has been deferred three weeks running.

Olma handles the rest. The journal mode structures what was said into wins, problems, blockers and needs, and files it against the company.

Local first, because of what people say

This only works if founders are candid, and candour depends entirely on where the audio goes.

That was already settled by how Olma was built. Transcription and rewrite both run on the device, so the recording of a founder describing a co-founder problem never travels. What leaves the machine is the structured entry, encrypted, to the fund's own store. The raw audio stays where it was spoken.

Building the dictation tool local-first was a preference. For this it became the reason the product is usable at all.

Turning entries into work for the fund

An honest journal is only worth collecting if someone acts on it, so the last stage is triage.

Each entry is read against what the fund actually has: its capabilities, its network, its CRM, and the portfolio around it. A blocker is not just logged, it is matched. A company stuck on enterprise security review becomes an introduction to the fund partner who has been through it twice. A company describing a hiring problem becomes a candidate the fund has met. A recurring technical problem becomes a piece of the fund's own tooling shared across companies.

The output is not a dashboard. It is a list of things the fund can do this week, each attached to the company that needs it and the reason it came up.

Where it is now

Olma ships as a dictation app today. The fund layer is the direction it is being taken in, and it is the reason the modes system exists in the shape it does.

The bet underneath it is simple. Funds do not have a reporting problem, they have a candour problem, and candour is a design question long before it is a data one.