Today we are introducing Wonderloom, a new product direction at Vibe Coding Agency. Wonderloom is an adaptive, illustrated learning engine, built on a Rust event-sourced core that separates learner state, curriculum planning, narrative continuity, and safety policy from the language model that realizes each scene.
The idea is simple: a learner should not feel like they are using a chatbot or completing a worksheet. They should feel like they are inside a story that remembers them, adapts to them, and teaches through the plot. The system is designed so the LLM writes dialogue and proposes scene mutations, but deterministic validators and a supervisor loop approve every change to learner state, world facts, and safety boundaries.
Why we are building it
Most AI tutors today are chat wrappers. They are stateless, improvisation-first, and optimized for engagement rather than evidence of learning. They can explain a topic beautifully, but they do not necessarily know whether a learner has mastered it, what misconception is blocking them, or how to weave practice into a narrative the learner cares about.
Wonderloom is architected as a stateful educational world simulator with LLM-powered realization. The engine stores four separate memory classes: episodic, pedagogical, narrative, and semantic. It event-sources every interaction. It plans scenes with typed contracts. And it keeps the model as an interchangeable provider behind a capability-aware gateway, not the source of truth.
What the preview contains
The public preview is adult-facing only. Visitors must confirm they are 18 or older and agree to the hosting provider's data residency terms before entering the demo. The demo uses fictional learner profiles only: a space enthusiast, an animal lover, and a puzzle solver. Each profile generates a constrained scene that illustrates how the engine adapts vocabulary, math, and reasoning beats to the learner's stated interests and the story's current dramatic question.
No real user data is collected in the preview. There is no account creation, no voice or video capture, no external browsing, no social features, and no advertising. The point is to give adult learners, educators, and investors a concrete product moment before we open paid plans.
Safety by design
- User-owned accounts. Every account is owned by the adult user. Learner identity is a local nickname; real names are not required.
- Data minimization. Voice, video, social interaction, and external browsing are off by default.
- Explainable memory. You can inspect, correct, export, or delete episodic, pedagogical, narrative, and semantic memory.
- No hidden persuasion. We optimize for demonstrable learning, not engagement time or emotional dependency.
- Safe degradation. When inference, retrieval, or policy evaluation fails, the system falls back to bounded local material or halts.
The business model
Wonderloom will launch as a free adult-facing preview, followed by a paid personal plan with persistent story worlds, a weekly Learning Journal, export and deletion controls, and richer multimodal scenes. We are designing pricing around cost per completed learning session, not unlimited chat turns. The goal is to prove retention, learning outcomes, and inference margins before expanding subjects and audience segments.
What comes next
We are starting with one narrow wedge: adult self-directed learners, reading comprehension, vocabulary, applied arithmetic, and scientific reasoning through a serialized illustrated adventure. If the core loop retains, we will add persistent narrative worlds, audio, generated scene art, and a learner dashboard. Team and organizational training is a later channel, not the initial launch.
If you are an adult learner, educator, or investor who wants to follow the build, try the Wonderloom preview and read the User Guide & Data Policy. We will publish updates as the beta takes shape.