Research should compound,
not evaporate
An umwelt is the slice of the world an organism can actually perceive. Most product teams have a very small one — bounded by whichever study someone remembers. We are trying to widen it.
The archive that can't answer
Teams run interviews constantly. They transcribe them, analyse them, present them, and then — almost universally — stop. The recordings go to a drive, the findings go to a deck, and within a quarter the only working index of what was learned is a person's memory.
That is a strange outcome for an industry that takes evidence seriously. The evidence is all still there. It is simply unreachable, because nothing in the stack connects a question someone has today to a sentence a participant said eight months ago.
The obvious fix — point a language model at the transcripts and ask it things — fails on the thing researchers care most about. It produces fluent claims with no provenance, and a claim you cannot check is worse than no claim at all, because it is more persuasive.
So Umwelt is built inside out from provenance. Atoms carry their speaker and timestamp. Insights carry their atoms. Themes carry their participant counts. Recommendations carry the whole chain. The structure is the product; the prose is just how you read it.
Augment the labour, not the judgement
What we automate
Transcription. Reading everything. Lifting observations. Matching evidence to patterns across studies you have not opened in a year. This is the work that scales badly and rewards nobody.
What we don't
Deciding what a finding means, which contradiction matters, what to recommend, and what should be said out loud to a stakeholder. That is the job, and it is not a bottleneck to be removed.
How we keep the line
Model output enters a review queue, not the corpus. It stays visibly proposed until a person accepts it. There is no path by which a generated claim becomes an official finding on its own.
Four commitments
Provenance over prose
A claim that cannot be opened is a rumour with good typography. Everything in Umwelt keeps the evidence it came from.
The human decides
The system proposes. A researcher disposes. Nothing enters the shared corpus because a model was confident about it.
Answers in seconds
If asking the corpus is slower than asking a colleague, nobody will ask the corpus. Speed is a correctness requirement, not a nice-to-have.
Participant data is sacred
People told you things in confidence. That obligation travels with the data through every system we put it in.
A small team, building in the open with users
We sit at the intersection of product management and UX research, and we built the tool we kept wishing existed. Umwelt is in private beta: we onboard teams personally, watch how the product holds up against a real corpus rather than a demo one, and ship against what breaks.
If you run research and something here sounds either exactly right or obviously wrong, we'd like to hear it. Both are useful this early.
Research tools that respect your expertise
Move faster without giving up rigour, and without handing your conclusions to something that can't show its work.