Wonda let creators build advanced immersive learning experiences in 3D and 360. Genuinely powerful, and genuinely out of reach for the teachers and learning designers it was meant to serve: anyone without prior immersive-media experience hit a wall before they started. My question as product manager was how to let a whole new category of user create something meaningful.

The answer was conversational AI avatars for language practice, manager training, interview rehearsal and patient simulation, with no 3D setup required. I shaped the roadmap, wrote the user stories, and produced the wireframes and interface design for this new layer of the platform.

Most of the work was designing around the model's real limits. GPT-3.5 had just launched and hallucination was severe, so the AI couldn't be positioned as an expert. A doctor or a history teacher was simply not viable. What it did exceptionally well was hold a natural conversation. Scoping the product to that strength, and refusing the use cases it couldn't support, is what made the feature trustworthy enough to ship.

Two design decisions carried the most weight.

Assessment that stays honest. Learners needed a report after each conversation, not just the conversation itself. Rather than let the AI produce generic feedback, we gave teachers control over their own evaluation criteria, and I separated the character prompt from the assessment prompt so the two can't interact. A learner can't talk the character into awarding them a good score.

Pricing that survived the token economics. Conversation costs are nothing like serving web content, which broke the existing free plan of three experiences with unlimited use. To keep a free tier at all, I introduced a message limit calibrated by trial and error: enough to talk with a character and generate a real report, not enough to run an actual exercise. Finding that line was a product decision as much as a commercial one.

The feature launched in beta with North Carolina State University for management training on cultural differences, alongside Harvard for language learning and the University of Sheffield for social care. The beta succeeded and AI characters are now a standard part of the Wonda platform.