Current build
What exists today
Example manual entry
Working in code
- Speech recognition, a local Qwen model, and spoken replies on Windows
- A private browser connection for a Mac or phone
- A fixed manual that returns a stored answer with its exact source
- Deterministic tests for known, missing, conflicting, and excluded questions
Not tested yet
- A real house manual or ordinary daily questions
- The Galaxy’s battery, microphone, browser, and reconnect behavior
- Whether the intended user prefers this to a pinned note
- Whether every device and service in the complete path leaves no logs
How it works
One button, one private route.
The phone is a microphone and screen, not the computer doing the work. A private Tailscale connection carries the request to the existing Windows PC. The browser has no tool or home-control path.
OpenWhispr’s Parakeet model handles speech recognition, Qwen runs through Ollama for general answers, and Kokoro reads the reply aloud. A small local file handles stored house facts before the general model answers.
Parts
Use the hardware already in the house.
Runs speech recognition, Qwen, the manual lookup, and speech output.
Planned push-to-talk interface. Its real device behavior is still untested.
Keeps the browser page off the public internet and checks the private connection.
Ignored by Git, read-only at runtime, and limited to low-stakes facts.
Local speech-to-text, general answers, and text-to-speech.
No new hardware is required for the first trial. That is a plan, not a measured total.
Answer rules
House facts and general answers don’t get the same label.
House manual
Stored answer. Exact source.
A house answer must come from the fixed manual and show its source, location, and checked date. Missing or conflicting entries return “Unknown.”
Source · location · checked date
General question
Local model. Not verified.
A familiar question such as “Who won the 2022 World Cup?” can go to Qwen. It is clearly marked as model output, with no promise that it is current or correct.
No live web · may be wrong or stale
Alternatives
Most of the ingredients already exist.
Samsung already lets a phone launch Gemini from the side button. Home Assistant Assist offers private voice on owned hardware. Open WebUI supports a phone microphone and local speech recognition. AnythingLLM and PrivateGPT answer from local documents, while HomeBox tracks household records.
That means “local AI on an old phone” is not a meaningful advantage. The only idea worth testing is the tighter contract: exact local sources for house answers, an honest “Unknown,” no control path, and no persisted conversation database.
14-day trial
It has to beat a pinned note.
Put 15–20 real, low-stakes facts in both the manual and a pinned note. Add five deliberately missing or conflicting questions. For two weeks, record which option the intended user reaches for during ordinary retrieval moments.
- Known answers
- Every answer is supported by the displayed source.
- Missing or conflicting
- Every case returns “Unknown.”
- Warm response
- Speech starts within 6 seconds on 9 of 10 turns.
- Actual use
- House AI is chosen over the pinned note most of the time.
- Upkeep
- The manual takes no more than 15 minutes a week to maintain.
Stop on one unsupported house answer, one missed conflict, any mutation or control path, persisted conversation data, nonlocal content processing, unsafe phone behavior, repeated preference for the note, or too much upkeep.
Limits
Deliberately small and low stakes.
- No web search, arbitrary file access, document upload, tools, or home control.
- No gas, electrical, fire, emergency, security, credential, or shutoff guidance.
- General model answers can be wrong or stale and are never house-manual evidence.
- The source code is private because it contains deployment details; the real manual is ignored and will not be published.
- The application has no transcript database, but complete device and service logging still needs live verification.
The next honest milestone is not another feature. It is the first real manual, the Galaxy check, and the two-week comparison.