What is Somia? The architecture of a conversational AI system for everyday emotional wellbeing

Somia™ is a modular conversational artificial intelligence system for everyday emotional wellbeing, developed by yeshcube. Its architecture, common to every solution that integrates it, is organized into five planes: experience orchestration, content, data, analysis and validation. The system is deployed in commercially available products, both our own and third parties’, from acoustic booths to relaxation masks and voice applications.

ERL status

The ERL scale (Evidence Readiness Level) measures the evidence maturity of a solution across six levels, from the initial hypothesis (ERL-0) to consolidated impact (ERL-5). Somia holds level ERL-3, initial transfer.

The problem: emotional language in conversational systems

General-purpose conversational models process linguistic patterns through statistical analysis, without real emotional understanding. In the field of wellbeing that approach runs into two difficulties. The first is interpretation: emotional language is full of metaphor, venting and ambiguity that a generic model can read literally, producing responses that are inadequate or potentially harmful when a conversation touches on delicate ground. The second is deployment: a technology intended to support people in contexts as different as an office booth, a rest mask or a family application has to behave with the same safeguards in all of them.

Solving both at once takes more than a good language model: it takes an architecture that interprets context before responding and carries its rules over to every product that integrates it.

Design requirements

Five requirements follow from that problem, and they govern Somia’s design:

• Contextual interpretation: the system analyzes what kind of expression it is receiving before generating any response.
• Bounded scope: its action is limited to non-clinical wellbeing, with stopping rules for anything beyond that.
• Privacy by design: personal information stays under the control of whoever uses the system.
• Measurement of usefulness: aggregate use is measured in order to evaluate and improve the solutions.
• Transferability: one and the same core has to work with identical safeguards across heterogeneous products.

Three design principles accompany these requirements: audio first, zero screens where the context allows, and applications without gamification or intrusive notifications.

Architecture: five shared planes

The architecture answers those requirements with five planes shared by every integration. Orchestration decides which experience is offered, in what sequence and under what rules. Content brings together the sound experiences, guided conversations and exercises each solution can activate. The data plane manages personal memories and usage signals. Analysis turns those signals, always in aggregate form, into measurement and learning. Validation applies the evidence and integration criteria every solution has to meet.

Being modular separates what the core provides from what each product resolves. The core supplies the conversational models, the interaction protocols, the safety rules and the measurement mechanisms. The end product decides the interface, the sound system, the connectivity and the environment of use: the same architecture sustains a conversation in an acoustic booth and a sound experience in a relaxation mask.

How an interaction works

A session with the conversational layer follows a defined path. The person activates the microphone, which is always voluntary and can be turned off. The audio is processed in real time through a technology partner specializing in voice, and transcribed. A multi-layer instruction system then analyzes the context: it distinguishes between factual statements, metaphors, venting and possible risk signals. That classification determines the next step, which may be generating a response, adjusting the experience under way, or triggering the rules for stopping and redirecting toward professional resources. The response is synthesised into speech, with turn-taking managed. The session transcript is used only to personalize the experience, apply the safety rules and keep the service running.

yeshcube calls the technical capability behind this path algorithmic empathy: interpreting contextual signals and adjusting the response accordingly. The term describes a behavior of the system, not an emotional experience of the machine.

The system operates in two modes. Somia Sensory runs predefined sound experiences that work without conversation, aimed at relaxation, breathing, focus or rest. Somia Immersive adds the conversational layer described above, which adapts the sequence, the language and the flow of each session to the available context.

Data governance

Data follows two separate regimes. The first is aggregate: Somia Data is the measurement and intelligence infrastructure that yeshcube is developing to turn patterns of use, continuity and usefulness into dashboards, reports and improvement decisions. It is at ERL-0 and its technical architecture has yet to be specified. Its principles exclude the sale of personal data, advertising profiling and the exposure of individual conversations in client reports.

The second regime is personal and stricter. Personalization memories are organized into 18 blocks across four dimensions: self-knowledge, emotional wellbeing, personal development, and a fourth devoted to the person’s environment and connections. They are enabled only by the person’s decision and can be viewed, edited or deleted at any time. The architecture is local-first: information lives on the device by default and is only synchronized to the cloud when that option is expressly enabled.

Processing is governed by the GDPR and the LOPDGDD. Yeshcube Tech, SL acts as controller, providers operate as processors, and international transfers are subject to standard contractual clauses.

Somia ecosystem infrastructure

Three infrastructures extend the system and connect its applications. Somia Within establishes the integration, ethics and evidence criteria that first-party and third-party products must meet. Somia Atlas provides the territorial layer for locating available spaces, booths, applications and experiences. Somia Data structures aggregate measurement, outcome reporting and the learning needed to improve deployments.

Within enables the system to be integrated under shared safeguards, Atlas supports discovery and access, and Data helps explain how the network is used and evolves. Together, they turn isolated solutions into an ecosystem that is connected, measurable and transferable across products, organizations and territories.

The model is already working in commercial products: Somia Bloom (AI spatial sound for acoustic booths, compatible with models from any manufacturer), VVAVVE™ (relaxation mask with immersive audio) and Kiwikidoo™ (voice adventures for families), alongside several ambient devices in development. The details of each product are on the Somia site.

Evidence and validation status

Somia holds level ERL-3, initial transfer: the system has been transferred to products, our own and third parties’, that are already commercially available. Level ERL-4 (validated scaling) requires demonstrating scaling with studies across multiple contexts and independent review. It will be declared once that evidence is available.

Intended use and limitations

Intended use. Non-clinical emotional support for healthy adults: managing everyday emotions, relaxation, focus and preventive self-care.

Out of scope. Somia does not provide medical, psychological or psychiatric services, does not make diagnoses and does not replace professional care. In a crisis, the right route is the emergency services or a mental health professional.

Technical limitations. The models can make mistakes, miss nuances or interpret context only partially. The risk detection mechanisms can produce false positives and false negatives, do not perform clinical intervention and do not contact emergency services. Faced with signals beyond its reach, the system stops the interaction and points toward professional resources.

Users’ rights. To know that they are interacting with a machine, to control and delete their memories, to object to uses of their data beyond the service, and to obtain human review of decisions they consider significant.

System documents

• AI Transparency Commitment: the nature of the system, memories, risk detection and its limits. Next review: November 2026.
• AI User Agreement: rights and responsibilities in interacting with the AI.
• Privacy policy: data processing, processors and international transfers.