When a student asks a question, Knowunity's system analyzes the query through multiple lenses: intent, complexity level, and subject area. The system then retrieves relevant content from its database and selects the appropriate system prompt and model size.

"We have specific system prompts for each intent," Hild explains. "One that explains how we solve math problems, another for translations, another for essay help." For complex mathematical problems, they built a specialized "Math Code Execution Engine" that prevents calculation errors by converting problems into executable code before returning natural language explanations.

The platform's infrastructure evolved with its ambitions. Starting with a simple monolith and database, Knowunity quickly recognized data as its competitive advantage. "BigQuery became one of our most important Google services," Hild notes. "We store all our data there – backend data, ML data, analysis data."

Knowunity doesn't rely on a single model approach. Instead, they strategically deploy different Gemini variants based on specific use cases. "We use almost all Gemini models for different use cases," Hild explains. "If it's a difficult math problem, we use Gemini 2.5 Pro. If it's a straightforward question, Gemini 2.0 Flash is sufficient." This optimization balances quality, cost, and latency, crucial factors when serving millions of students who expect instant responses.

The platform conducts extensive A/B testing to optimize model selection. "We can always see what impact a model switch has for a specific intent on student retention and engagement, as well as on feedback we collect directly in the companion," Hild notes. "We've learned a lot over time about which model to use, which system prompts to use, which tool calls to include, and how to adapt the UI so students can learn as effectively as possible with the content."

The platform implements multiple safeguards through a three-step content moderation process: machine learning models for initial filtering, manual team review for edge cases, and ongoing community feedback. "Content quality has never been a huge issue because we mainly operate at the school level with foundational knowledge, where accuracy standards are well-established and clear," Hild notes.

The results speak for themselves. Knowunity has grown to serve 20 million students across 17 countries, with concrete learning improvements. "Through surveys, we see that when a student has been with us for two weeks, their grades improve by 0.5 points on a scale of 1 to 6," Hild reports.

Google Cloud's global infrastructure has enabled this expansion. "The global infrastructure was a big plus point for us," Hild explains. "We can assume that Google has a data center locally in almost every country, so we're relatively close to our customers."

This proximity translates into improved user experience and higher retention. The peer-to-peer content model has created a virtuous cycle where students create content that helps others in their region. At the same time, Knowunity's AI systems learn to provide better personalized recommendations.

Schools have taken notice. "We regularly get inquiries from schools that would like to buy the app for their entire school," Hild reports. Rather than seeing Knowunity as a competitor, educators are increasingly viewing it as a valuable supplement to traditional instruction.

The platform has also created an unexpected revenue stream by connecting students with universities and apprenticeship programs. "Many students don't know what to do after school," Hild explains. "On the other hand, there are many universities desperately looking for students, and apprenticeship positions looking for trainees. By bringing these two parties together, we were able to build a second, quite substantial revenue stream."

The potential to transform education systems is enormous. "I think there's potential for education systems like Germany's to increasingly rely on such technologies to raise educational standards fairly for everyone," Hild observes. The vision of every student having a personal learning companion from fifth grade through graduation could fundamentally democratize access to personalized education.

Technical innovations continue to drive toward this future. "Dynamic visualizations are currently a huge internal topic for us, answering with visual content in addition to text," Hild explains. "We've made good progress, but especially when it comes to showing students dynamic visualizations, the current state of AI isn't quite at the level where we can work well with it yet."

The platform is also exploring expansion beyond traditional school education. "Education doesn't stop after the high school diploma," Hild notes. "Our goal is to accompany students for as long as possible, so naturally, university is the next logical step for us."