Most discovery tools surface what's already trending — which is the opposite of what A&R is for. Nomad A&R is a sub-genre embedding space trained on the audio itself, calibrated against the working judgement of A&R teams who've actually signed acts, and built to be disagreed with. The model proposes; the team decides; both get better.
If your discovery tool ranks by current streams, current playlists, or current social momentum, it will only ever find acts you already know about. By definition. The point of A&R is to find what no one has found — to hear the act that the playlist editor hasn't heard yet, that the algorithm hasn't surfaced, that the manager hasn't decided to pitch.
Nomad A&R is built specifically not to do that. The embedding is trained on the audio — on what an act sounds like, not what it charts like. Two acts are close in the embedding because they sound close to a careful listener, not because they share a playlist or a follower base.
Trained on 78M tracks using a contrastive objective: tracks that are confirmed sonically related (by working A&R judgement, not by playlist co-occurrence) are pulled together; unrelated tracks are pushed apart. No streaming signal in the loss function.
We calibrated the embedding against signings made by twelve working A&R teams across 2023–2025. The model recovers their actual signings inside its top-2% recommendations with 71% recall. Recall doesn't measure success; it measures that the embedding is not blind to the kind of judgement A&R actually exercises.
Every recommendation surface comes with three audio references and the implicit framing the model used. When your team disagrees with the framing, that disagreement is logged and fed back into a re-ranking layer specific to your roster's taste.
Over six months of use, the model develops a roster-local taste profile that reflects your team's actual decisions. The base embedding stays general; the re-ranking layer learns the specific way you hear.
A typical week: every A&R member gets a Friday queue of 12–20 candidates the embedding believes match their current focus — a sub-genre they're scouting, a roster gap they've flagged, or a sound they've been calibrating against. Each candidate comes with three reference tracks ("we think this is close to these three"), the implicit framing the model used, and a short audio sample.
The team listens. They mark each candidate as in-scope, out-of-scope, or interesting-for-different-reasons. Every mark is feedback. Every disagreement with the framing is feedback.
Over six months, the Friday queue becomes the highest-signal A&R surface most teams have ever had. Not because the model is smarter than the team — because the model is doing the part the team didn't have time for, and surfacing the candidates worth a closer listen.
Truly genre-creating acts are by definition outside any existing embedding space. The model will never find them. That work still belongs to humans. We are not pretending otherwise.
The model surfaces candidates worth a longer listen. The signing decision involves the artist's writing, the artist's commercial trajectory, the artist's manager, the artist's plan for their own career — and we model none of those. We never will.
Embeddings have edge cases — production techniques that fool the timbre channel, lyric-dense genres that surface badly without an English-first listener, traditional music outside the training distribution. Every confidence band on a Nomad A&R recommendation is real and worth reading.
We onboard at most 4 new teams per quarter. The waitlist is short; the calibration period is real. If you run A&R at a label, publishing co. or management firm and want to talk to us about access, the door is open.