KidsAI Buddy
Kid-safe AI tutor used daily by seven kids — fully local inference, fully local TTS.
The AI part
Socratic-method system prompts that ask guiding questions instead of giving answers, wrapped in a four-layer safety pipeline (profanity filter → jailbreak detector → homework-integrity check → topic boundary) — all running on a Proxmox VM, no kid utterance leaves the home network.
Stack
Why I built this
My kids wanted to talk to an AI. I wasn’t going to hand them ChatGPT. So I built one whose entire job is to be kind, curious, and incapable of doing their homework for them.
How it works
- Socratic by design. The system prompt forbids direct answers on homework; the model asks the kid what they think first, then nudges. A separate “homework-integrity” classifier flags suspicious requests for the parent dashboard.
- Four-layer safety. Profanity filter → jailbreak detection → homework-integrity check → topic boundary. Layers run before and after generation so a model that drifts gets caught.
- Voice mode. Browser STT + barge-in (the kid can interrupt mid-sentence). Speech synthesized locally by Kokoro-82M — no audio touches the cloud.
- Gamification. XP, levels, streaks, badges — but transparent. The kids can see exactly how the points work; no dark patterns.
- Parent admin. Conversation review, safety-flag triage, per-kid difficulty selector (Light / Medium / Hard).
- 15 themes, 126 emojis. Because making it theirs matters more than I expected.
What it took
- A FastAPI service plus a Kokoro TTS sidecar in its own container — model updates ship without taking the chat down.
- Cloudflare Tunnel for outbound-only exposure (no port-forward, no public IP).
- Welcome tour onboarding so a kid can use it without an adult walking through.
- Detailed audit log behind the parent dashboard.
What I learned
Safety in kids’ AI isn’t a single classifier with a high recall number. It’s a layered system where the post-generation check catches the things the pre-generation check missed, and where the parent dashboard makes the unavoidable failures visible instead of silent.