Products · Serena

Read emotion.
See the working.

Serena is a symbolic affective AI that reads emotion in language and exposes its full six-layer trace for every turn. White-box. Calm. Yours to audit.

Open SerenaTry the playground

Features

Six things you can audit.

01

12-emotion vector, no training data

Reads affect directly from ARPAbet phoneme statistics. Every channel is interpretable; no learned weights to attribute.

02

7-dimensional lifespan model

STAB, OPEN, AGREE, ASSERT, SENSE, LEARN, plus accumulated experience. The agent's developmental curve is part of its state.

03

Six-layer trace per turn

Phonetics → text features → affect → developmental modulation → cognition → generation. Every layer inspectable.

04

Stateful, on purpose

Per-user emotional state persists across turns within a session. Reset is one click; nothing is hidden.

05

Custom system instructions

A CLAUDE.md-style file you can edit. Prepended to every analyse and chat call. Yours.

06

API and Studio included

API keys, agents, webhooks, a playground that doesn't burn your quota — at studio.serenachat.com.

How it works

One trace, six layers.

Each turn moves through the pipeline below in order. Every layer's output is captured and visible in the trace.

L0Tokenise + POS-tag
L1Phoneme features
L2Text features
L37D + 12D affect
L4Cognition + memory
L5Generation

See a real trace at serenachat.com/app/analyze.

FAQ

Honest answers.

Is Serena trained on my data?

No. Serena is symbolic; there is no training step that learns from your inputs. Your text travels through the pipeline and is discarded — except for any traces you choose to save in your account.

How is this different from running a sentiment model?

A sentiment model gives you a label and a confidence score. Serena gives you the full pipeline state — phonetic features, two affect vectors, developmental context, and the decisions made at each step. You can audit how the answer came about, line by line.

Can I run it on-premise?

Enterprise customers can. The default deployment is hosted in EU/UK regions on Google Cloud Run; an on-prem option is available on request — see the Business page.

What languages does it support?

English today (because the phoneme analyser is built on ARPAbet). Other Latin-script languages are technically possible with a phoneme map; we'll publish each one as a release rather than overpromising.

Does Serena make clinical claims?

No. Serena reads affect from the surface of language. It is not a diagnostic tool, and we don't sell it to anyone who would treat it as one.

Try it now.

The Analyze tool is free and open. No card needed.