Blip
A dedicated AI learning console for kids. Voice-first, TV-connected, privacy-respecting. Wake word: "Hey Blip."
FOR PARENTS
What Blip does
Learning modes, activities, hardware, and privacy details.
FOR BUILDERS
Collaborate
Open roles in design, engineering, and education.
FOR INVESTORS
The pitch
Market, unit economics, business model, and the ask.
What it is
Blip is a mini PC tucked behind the living room TV. My kids say "Hey Blip" and it wakes up, listens, and responds. It runs spelling bees, math practice, trivia, interactive stories, and speech practice. When the TV is off, it works audio-only. When the TV is on, it shows visual feedback and content.
Why I built it
Voice assistants give shallow answers and are ad-adjacent. Tablets are dopamine traps with a thin educational veneer. Educational apps are isolated — no real conversation. I wanted something that felt like talking to a smart, patient tutor, not a product.
Privacy mattered too. Speech-to-text runs locally on the device. My kids' voices never leave the house. The only data that reaches the cloud is transcribed text — and even that is routed to the local blip-edu model for most structured activities.
How it's built
- Hardware: Beelink SER5 MAX (Ryzen 5, 16GB) · Anker PowerConf S330 USB speakerphone · HDMI to existing TV
- Wake word: openWakeWord — "Hey Blip" — runs fully local, Apache 2.0
- Speech-to-text: faster-whisper (small.en, CUDA), local, English only
- Text-to-speech: edge-tts (current) → Piper local (Phase 3)
- Intelligence: Hybrid — Claude Sonnet 4.6 (API) for open-ended conversation, storytelling, and emotional support + blip-edu local LLM (Qwen2.5-7B LoRA, Ollama) for spelling, math drills, trivia, and greetings
- blip-edu: Custom LoRA fine-tune on 14,000 synthetic kid-tutor examples. Trained on RTX 4090. Three variants: base (Qwen2.5-7B-Instruct), coder base (Qwen2.5-Coder-7B), and blip-edu v1. Model selection validated by benchmark against 12 other models.
- TV UI: Electron app, fullscreen, large text, kid-friendly design
Current status
- ✓ Phase 1: Audio core — wake word, STT, TTS, event bus
- ✓ Phase 2: Claude integration + orchestrator + hybrid LLM routing
- → Phase 3: Electron TV UI + WebSocket bridge (active build)
- Phase 4: Learning modules — speech therapy ✓, literacy spec done, spelling/math/storytelling/trivia planned
- Phase 5: Kiosk mode, parental controls, polish
- Phase 6: Block coding UI, multi-child profiles, offline mode
Build log
I write about this as I go in the Building & Tinkering section.