Skip to content
Iterus

Innea · live

Innea: AI Support Between Therapy Sessions

Innea is an Iterus product: Czech-language AI support for people who need someone "between sessions" or don't have a therapist yet. Text and voice chat, mood tracking and a journal, plus a crisis layer designed to recognise risk signals and point to the 116 123 helpline. It is not a health service — it is a bridge. Live since spring 2026.

What problem does Innea solve?

Between two therapy sessions there is a week, sometimes two — and what weighs on a person does not wait for the appointment. Whoever has no therapist yet has no such rhythm at all. Innea builds a safe space into that gap: a conversation in Czech that remembers context, tracks mood and flags themes that keep coming back.

The boundary is essential. Innea is not a health service and does not pretend to be one — it is support between sessions, not a replacement for them. And when a conversation points to a crisis, the crisis layer refers to the 116 123 helpline and hands the case to a queue handled by a human.

What did we build?

  • AI chat in Czech guided by a therapeutic conversation framework, with three-layer memory (the current conversation, a semantic memory of earlier conversations, long-term memory) on the Claude API and automatic failover to another provider during an outage.
  • Voice chat — speech recognition and synthesis (Deepgram, ElevenLabs), falling back to text for sensitive questions.
  • Mood tracking and a journal with short- and long-term trends and AI detection of themes that keep coming back.
  • Crisis layer: risk-signal detection tuned to Czech → assessment by a language model → content safety filter → an administrator queue with a tracked response time. Protection against alert overload makes sure a full intervention is not triggered by every sentence.
  • Subscription and a free tier through Stripe — the crisis layer works regardless of the free account's limit.
  • Compliance: GDPR, a completed DPIA, data export and deletion on request, and an EU AI Act transparency layer — the user knows they are talking to an AI.

The numbers

Innea and Innea Pro share one codebase, so the numbers below describe the whole platform — from the first commit in October 2025 to the state measured in September 2026.

unit tests (Vitest)
6,115
end-to-end tests (Playwright)
478
API endpoints
572
database tables
243
RLS policies (row-level data isolation)
636
database migrations
349
merged pull requests
590
custom CI guards (secrets, PII, security)
10

We don't publish operational numbers yet (accounts, messages, sessions): these are the first few dozen users and the figures would not mean much yet. We'll add them once they do.

How did we work?

We work AI-native: delivery is accelerated with Claude Code and Codex, and quality is secured by automated tests, independent code review and specialists on demand.

Innea shows what that sentence means in practice.

The human — the project's technical lead — made the architectural decisions: Next.js App Router, Supabase and Postgres with each user's data strictly separated in the database itself (row-level security), a data model for two connected applications with graduated consents. He set the security model — crisis gates that intervene when in doubt, audits of database permissions, a rule for every irreversible (destructive) database change — and the process: tests written before the code (TDD), every change as a separate pull request with automated tests and AI review, merged only after human approval. He laid down the rules the AI must never break: no advice on medication, no reality-testing of psychoses, always a pointer to the crisis line. And he verified the result against reality — by hand in the browser and by checking production, which repeatedly surfaced defects that had passed both the tests and the AI review.

Claude Code wrote production code, database migrations and security policies to specification, tests ahead of the implementation, documentation and refactors; it prepared changes as pull requests and delegated adversarial review to other models.

An honest footnote: on Innea, roughly two out of three AI-review findings do not survive verification against the code. That is exactly why a human decides — and why it matters who leads the work.

What's next?

Innea is one of two products on the same core; the other is Innea Pro for therapy practices. The app lives at app.innea.cz.

We use the same approach — the Claude API with protective layers, data in the EU, tests and review — on client projects too: see AI integration, or simply get in touch.