Loading page
Datapoints, the shared timeline, and how panels stay in sync.
Every panel in AudioLens — Saliency, Attention, Embeddings, Perturbation, Diagnostics, Fairness, Layer Probes, J-Lens, Dataset EDA — reads from and writes to the same underlying job data. Instead of separate tools that each need their own upload and their own model run, AudioLens keeps one timeline per audio clip and model, and every analysis you launch attaches its result to that timeline.
A datapoint is one audio clip paired with one model. Everything you do —
run a prediction, compute saliency, probe a layer, slice by fairness group —
is an operation on that datapoint. A job (POST /jobs) runs one operation
against one datapoint and returns a job_id; the UI polls GET /jobs/{id}
until it's done, then reads GET /jobs/{id}/result.
Because every panel is reading results scoped to the same datapoint, switching
from the Saliency tab to the Attention tab or the Embeddings tab doesn't
require re-uploading audio or re-selecting a model — the frontend already
knows which clip and which model you're looking at. Panels that depend on
each other's output build on this directly: for example, the Embeddings tab's
clustering (HDBSCAN) and nearest-neighbour retrieval reuse the same
embedding space that the Dataset EDA tab visualizes, and J-Lens Lab
(/j-lens) fits a lens over the decoder layers that the J-Lens panel then
applies to a specific datapoint.
The operations underneath the panels are dispatched to three Celery queues so that fast, cheap work doesn't wait behind slow, heavy work:
| Queue | Operations |
|---|---|
gpu-fast | prediction, embedding, hidden states, layer probing |
gpu-large | attention, saliency, faithfulness, Jacobian Lens |
cpu | perturbation, audio features, linguistic-vs-acoustic sweep, fairness, cleanup |
AudioLens has no user accounts. The first time you visit, the API sets an
anonymous, HttpOnly sid cookie with a 24-hour TTL. Every resource you
create — uploaded audio, jobs, custom models, custom datasets — is scoped to
that session cookie, not to an identity. There are no roles and no
multi-tenancy: a session either has access to what it created, or it
doesn't. When a session's data expires after 24 hours, it's simply gone —
there's nothing to migrate or reclaim.