
We built TurtleTalk around a lightweight, privacy-first AI architecture designed specifically for emotionally sensitive conversations with children. Instead of relying entirely on large cloud-hosted models, we fine-tuned Alibaba Cloud’s Qwen3-0.6B using 4-bit QLoRA optimization, enabling efficient on-device inference while maintaining strong conversational quality. The training pipeline was implemented through the Unsloth framework inside a containerized environment deployed on AMD Cloud infrastructure for accelerated experimentation and reproducibility. Our fine-tuning dataset and prompting strategy were built around established Social Emotional Learning (SEL) principles, focusing on emotional reflection, empathetic dialogue, self-awareness, emotional regulation, and curiosity-driven questioning for children. We optimized the model to support TurtleTalk’s three core interaction modes: Reflection Helper, Venting, and Big Questions. During training, the loss curve consistently decreased and later stabilized, indicating convergence toward reliable conversational behavior without severe overfitting. The resulting model checkpoint was successfully published to the Hugging Face Hub for versioning, reproducibility, and future community collaboration. A core technical decision behind TurtleTalk is local inference. By running the fine-tuned model directly on-device whenever possible, we significantly reduce response latency, creating a more natural conversational experience for children during emotionally sensitive moments. Local execution also minimizes the need to transmit personal conversations to external servers, providing stronger privacy guarantees and improving trust for both children and parents. This architecture positions TurtleTalk as an emotionally aware AI companion that is not only lightweight and responsive, but also fundamentally designed with safety, privacy, and child-centered interaction in mind.
10 May 2026

"A second grader gets pushed at recess. She doesn't tell her teacher — she's embarrassed. She doesn't tell her parents — she doesn't want to worry them. By the time an adult notices, it's been three weeks. This happens in every school, every day. Tattle Turtle exists so no kid carries that alone. Tammy the Tattle Turtle is an AI emotional support companion running on a simulated Reachy Mini robot. Students walk up and talk to Tammy through voice. She listens, validates, and asks one gentle question at a time — max 15 words, non-leading language, strict boundaries between emotional triage and treatment. What makes Tattle Turtle different is what happens beneath the conversation. Every exchange is classified in real time into GREEN, YELLOW, or RED urgency. A bad grade vent stays GREEN — private. Recess exclusion mentioned three times this week? YELLOW — a pattern surfaces on the teacher dashboard that no human could track across 25 students. A student mentions being hit? Immediate RED alert — timestamp, summary, and next steps pushed to the teacher. The system comes to them when it matters. We built this on three sponsor technologies. Google DeepMind's Gemini API powers the conversational engine with structured JSON for severity and emotion tags. Reachy Mini's SDK provides robot simulation through MuJoCo with expressive head movements and audio I/O. Hugging Face Spaces serves as the deployment layer — one-click installable on any Reachy Mini in any classroom. Tammy's prompt engineering uses a layered 5-step framework ensuring she never crosses clinical boundaries, never suggests emotions to students, and never stores identifiable data. Privacy isn't a feature — it's a constraint baked into every layer. Tattle Turtle fills the gap between a child's worst moment and an adult's awareness. One robot. Every classroom. No kid left unheard."
15 Feb 2026