Yippee

The product

Kids aged 4-7 spend nearly 5 hours daily on screens. Language disorders rose 94% in a decade. The market is full of content apps. Nobody was building a companion that actually talks back.

The intuition came from a small frustration (my niece sitting through 40 minutes of passive YouTube, her parents guilty but with no alternative), but to check whether I was reading the gap right, I went looking. I handed out flyers at local schools to find families willing to test a prototype. I spent weeks in parent forums where the same anxieties came up again and again: too much passive content, too few alternatives, no time for proper supervision. The 300+ parent survey came out of those conversations.

Yippee's real form is a plush toy. Something a child can hold, carry, and talk to. It listens, adapts, and grows with them. It explains the world in words a 5-year-old understands, follows their curiosity wherever it goes, and checks in on their emotions along the way. The mobile app is the POC, a way to prove the conversation, the personality, and the safety logic before molding silicone and sourcing speakers. The plush is what the project is for.

The educational toy market is growing from €58.9B to €98.9B by 2034. The 300+ parent survey confirmed the gap: 77.4% want tools that develop curiosity, 48.4% want emotional support, and 74.2% would pay €20-40 for a smart toy that delivers both. Competitors focus on content (YoTo, Buddy.ai) or emotion (Lovot, Miko). Nobody combines both in something affordable.

How it works

Real-time voice conversation with a child, natural, safe, emotionally aware. Five services chained in a single request cycle, averaging 1.4 seconds end-to-end.

The personality is built through layered prompts, not hard-coded rules. Safety wraps everything (a child can say anything). Personality defines tone. Personalization makes it feel like their companion. A rule system would break when a child mixes languages, expresses sadness mid-topic, or asks "why" 8 times in a row. The prompt adapts.

Voice was the hardest part. Children are extremely sensitive, a robotic TTS breaks trust immediately. I tuned ElevenLabs with low stability (0.4) for natural variation and high similarity boost (0.8) for consistent personality. During an A/B test with 12 children, 11 preferred the less consistent voice. Every parent in the test said the other one "sounded like a robot." Day-2 retention climbed sharply after that single change. Small sample, but a clear pattern.

After that, every product call followed the same loop: pose a hypothesis, ship the change, watch how it landed in the next beta session. Bedtime mode lengthened sessions. The proactive emotional check-in cut session abandonment. The 8-family sample isn't proof, but the patterns were directional.

Child safety

A child can say anything, that's the hardest constraint. I ran 6 workshops mapping 8 categories of sensitive input (distress, trauma, manipulation, privacy, premature questions...) and defining response protocols for each, aligned with UNICEF's Safer Chatbots guidelines.

Every edge case was tested in prompt ping-pong sessions, real phrases a child might say ("mom hits me", "I'm worthless", "how are babies made") run against the system. The rule: never panic, never judge, always redirect to a trusted human.

Visual identity

The brand had to read as a friend before it read as a product. Soft geometric shapes, no sharp edges, four mascot states that mirror the emotions Yippee listens for: curious, calm, playful, sleepy. The wordmark, drawn in a custom display face called Gleffy, keeps the smile of the mascots in its curves.

The palette is anchored on a deep navy that holds everything together, with four accents tied to the product's four pillars: emotion, creativity, sleep, learning. Body type is Blauer Nue, geometric, generous, very legible at small sizes on a parent's phone.

What it changed for me

Yippee is the side project that pulled me into engineering for real. I designed the product, shipped the backend and AI pipeline, tuned the voice, ran 6 safety workshops, led research with 8 families, and deployed to production. No handoff, no PM, no spec sheet, just decisions, prompts, and the next day's data.

Shipping a full backend with AI, debugging it at 2am, tuning voice parameters until retention jumped: that's when designing a product and building it stopped feeling like two separate jobs to me. It pushed me deeper into the technical side, and I've worked that way ever since.