← Blog Eugene Arutyunov October 6, 2026

Every Conversation Leaves a Fingerprint

The Velma conversation fingerprint is the house data visualization for conversations — the signature mark of the design language. Wherever a conversation shows up — in the product, in a report, in a demo — it leaves the same recognizable picture. Here it is in its fullest form, paired with the player:

Anatomy

The fingerprint is one component that flexes within fixed ranges. Every picture in this post is the same component — only the settings differ.

The X axis is a timeline that operates on fragments of speech — clips — rather than raw audio. The Y axis discretely splits the picture by speaker, one lane per voice. The standard encoding is color for detected emotions and a glyph for detected behaviors.

The information dial

How much a fingerprint tells is a dial. Bare clips only mark who spoke and when:

Detected emotions add color, and detected behaviors add a glyph on top:

One lane per voice

The number of speakers is 1, 2 or N. A single speaker reads as one strip:

A full room splits the same height into lanes:

Color is emotion

Clips are painted with the detected emotion. The colors come straight from the design system — one color per emotion in the model's vocabulary:

Every scale

The fingerprint speaks the same language at every size — that is what makes it an asset rather than a chart.

Icon scale

At the smallest scale the language compresses to a few bars. These are actual design-system icons:

Miniatures

Between the icon and the full view live miniatures — fingerprints in a table row, next to the conversations they belong to:

Or a plain flow of conversations, wrapping like text:

Full view

The ladder tops out at the full view — and there the fingerprint can also stand alone, as pure data, without the player chrome this post opened with:

Lanes beyond speakers

The Y axis can be rethought: lanes can be technologies instead of speakers. Here the same call is analyzed by two engines, one lane each, each with its own transcript variant. A clip that carries a signal blooms its color into the lane; a glyph with a count aggregates several signals:

The comparison can also isolate one selected signal: every flagged clip gets the same one-step highlight, and the glyph tells the case — a catch, or a red-marked false positive:

In the wild

The format scales past illustrations: the fingerprint can anchor a whole animated demo widget. Below, the same player drives Velma's fraud-call demo — the transcript, the signal feed and the risk meter all run off the timeline. Click anywhere on the plate to play:

Everything in this post is generated — every picture is the same renderer with different settings. The playground has the dials: set the room, flip the switches, export the result as data, or import a real conversation.

Try it: Velma Conversation Fingerprint.