Loading...
The AI infrastructure behind Gugu — built in Austria.
Behind Gugu runs a purpose-built orchestration and safety layer on top of generative AI. The platform creates personalized children's stories in text, image and audio — age-adaptive from 2 to 14+, multilingual, and built on strict safety-by-design. It has been running in production on the App Store and Google Play since June 30, 2026. This page explains what happens under the surface — and what we are researching.
Every story passes through a multi-stage pipeline:
Age level, selected characters, theme and learning goal are combined into a structured narrative context — the age-scaling engine calibrates vocabulary, sentence complexity and chapter size.
Inputs pass local filters and heuristics before they ever reach a language model — child-appropriate constraints are already part of the prompt.
Text, images and audio are generated in parallel across multiple AI providers; a job queue orchestrates the steps and streams finished chapters to the device immediately.
Generated content is AI-moderated after the fact — with conservative fail-safe behavior: when in doubt, content is not delivered.
Every generation is logged with tokens, assets and model choice — the basis for quality benchmarks and sustainable pricing.
Text, image and audio generation run across multiple AI providers (incl. Anthropic, OpenAI, Google). A custom orchestration layer picks the right model per task and balances quality against cost. Circuit breakers and waterfall fallbacks absorb outages of individual providers automatically — the platform keeps delivering even when a provider goes down.
Seven age levels from 2 to 14+ (2–4, 4–6, 6–8, 8–10, 10–12, 12–14, 14+): vocabulary, sentence complexity, chapter length and themes are calibrated per level. The same story idea is told fundamentally differently for a toddler than for a teenager — reproducibly and measurably, not by chance.
Child safety at Gugu is architecture, not an afterthought filter. Tier 1: local filters and heuristics before generation. Tier 2: child-appropriate constraints at the prompt level. Tier 3: AI-assisted moderation of generated content. Fail-safe behavior is conservative — when in doubt, content is not shown. Scientific validation of this approach is planned as a research cooperation: how do you guarantee high safety under generative uncertainty?
Characters should look and "sound" the same across chapters, stories and series — an open research problem for generative image models. Gugu uses a custom consistency pipeline for recurring characters so yesterday's favorite hero is recognizable tomorrow.
Today Gugu uses leading cloud TTS providers including voice cloning (parents can have stories narrated in their own voice). In development: our own German-language, child-friendly TTS model based on fine-tuning open models — with GDPR-compliant data anonymization in the training process and MOS evaluation as the quality benchmark. The goals: data sovereignty in Europe and long-term independence from US cloud providers.
A custom real-time sync architecture (WebSockets + client-side cache) streams stories to the device chapter by chapter as they are created. The child reads chapter 1 while chapter 3 is still being generated — no waiting-room feeling. Finished stories are available offline.
Generative AI for families has to be affordable. Every story is tracked at the token and asset level; model choice per task is a continuous optimization between quality and cost. In development: semantic caching, so similar requests reuse existing assets instead of regenerating them.
What we are working on — partly internally, partly in cooperations we are seeking with universities:
Gugu is developed by theconcept technologies FlexKapG in Styria and Vienna. We are continuously expanding our research and development activities in Austria and are seeking scientific cooperations with universities as well as funded R&D projects — from speech technology to learning research.
You do research on speech technology, AI safety for children or reading acquisition? You are a funding body, university or technology partner? We would love to talk.
Get in touchAn algorithm as a children's book author? How we orchestrate a gripping and flawless bedtime story in tenths of a second from 'Knight, Mars, and Astronaut'.
Artificial Intelligence is powerful – but it also carries risks. How the Gugu team built a strict filtering and guardrail system that guarantees parents full control.
Distance used to be a knockout criterion for grandparents reading bedtime stories. How Gugu's magical Premium Voice Cloning feature breaks down spatial distances.