Home systems / The companion / October 2026
Meet Berry.
A strawberry on a spare phone that helps the household and makes room for a little curiosity.
I wanted a friendly way for children to use the family system without needing to read a dashboard. Berry lives on a repurposed Pixel 2. Tap the strawberry, say something, hear what it understood, and confirm the change. BuryBoard displays the things it changes.
The character is playful. The rules underneath it are very deliberate.
01 / How it grew
A small helper, built in working slices.
I built this with AI agents, using separate implementation and review work. Each stage had to connect to the actual system before the next layer counted.
One household state
The companion and the TV board were connected to the same durable household records. Completing or undoing a task needed to mean the same thing on both screens, including after restarting the software. This was the foundation for everything else.
Voice joins the loop
Local transcription and speech made the phone a voice-first interface. Grocery and dinner requests and family votes could appear on the board after confirmation. I physically checked a request moving from the phone to the TV. Berry's learning shelf also expanded to 33 authored discoveries with pictures and two levels of explanation.
Make the invisible visible
The interaction review focused on listening, confirming, saving and getting stuck. Clear finish and cancel controls, picture-led proposals and visible confirmation playback replaced states that depended too much on reading. Grocery removal also gained a bounded, confirmation-required path for the selected profile's own requests.
A reader around the pictures
The original discoveries became a library: browse a world, open a topic, hear a passage, pause, replay and follow another picture. The younger reading band gets one idea per page; the older band gets a fuller passage. The same source material supports both.
A broader, reviewed pilot
A deployed content pack expanded the library to 120 topics across twelve worlds, with 360 reviewed sections and 109 approved media assets. Search could suggest reviewed questions. Automatic semantic answers stayed disabled because their evaluation did not meet the release gates.
02 / Character and interaction
Cute has to communicate.
Berry's drawing is an original SVG character: warm coral, pale seeds, green leaves, large eyes and small limbs. The app changes its expression and motion as it listens, thinks, speaks and asks for confirmation. Those states give a child another way to understand what is happening.
The more useful changes happened around the character. An early screen offered two similar-looking ways to stop recording, but one sent the speech and the other threw it away. For a child, that is a pretty expensive distinction to hide in a button.
Finishing became “I'm done.” Canceling became “Never mind.” A ring responds to the microphone, and a countdown shows the short recording window. Berry starts listening only after a tap.
A proposed change gets its own card: a picture for the destination, a visible action, the selected profile and the item. Yes fills while Berry reads the confirmation and becomes available when playback completes. No works while the question is still playing. A correction starts that confirmation over.
Recovery needed its own visual language, too. Unclear speech gets a puzzled response and a simple retry. An unavailable family computer makes Berry look as though it is resting. The explanation and reconnect control belong in a grown-up card. A cable problem should not look like the child is in trouble.
03 / Finding the voice
A voice needs an audition.
Local speech got the first working version talking. Later, I tried ElevenLabs' Flicker voice, heard Berry using it and rejected the sound. Ruby Roo replaced it. The service verified cloud synthesis with the selected voice, while retaining local speech as a fallback and keeping the existing usage limits.
That choice is part of the product. A voice name or a successful audio response does not establish that the character feels right on the phone. Hearing it in the actual interaction matters. So does keeping the system useful when the external voice service is unavailable. When cloud narration is used, the text being spoken goes to that provider; the reviewed learning content itself remains a local pack.
04 / The little encyclopedia
Pictures first. Sources underneath.
The goal is an explorable, narrated encyclopedia: choose something interesting, see it, hear an explanation and keep going. The deployed pilot covers space, Earth, animals, plants, machines, math, places, history, music, words, people and the body.
Topic pages offer narration, sentence highlighting and replay. A short trail makes it possible to return to recent topics. Pictures are identified honestly as reference images, diagrams or drawings, and the grown-up evidence view records where the wording and media came from.
The pilot includes source snapshots, literal supporting passages and media rights records. The sections received automated peer review by a separate drafting worker, with corrections retained. That is source-backed automated review, not expert certification or a substitute for a parent checking suitability.
Search did not earn permission to answer automatically.
The held-out semantic evaluation found the expected evidence in its top five results on 102 of 120 supported probes, or 85%. That missed the proposed 95% gate. Context handling also missed its gate. I kept automatic semantic answers disabled.
The released version can suggest existing reviewed questions for the child to choose. Selecting a question opens its reviewed passage. Search helps find the question; it does not get to invent the answer. Missing or broken optional search models leave the reader usable.
05 / Under the strawberry
One household, two surfaces.
Speak on the phone
The local host transcribes a short recording and interprets the supported request. Berry reads back a proposed change.
Confirm the exact change
Only a confirmed action reaches the shared household store. Saved operation receipts let the same request be retried without creating a duplicate.
Read it on the board
BuryBoard reads an allowed projection of that state. It does not maintain a competing copy of the lists. Adult-only text stays out of the child interface and public board projection.
The library is a separate, read-only branch. It retrieves reviewed reference content and narrates it. It has no authority to change household records. The friendly character can connect these experiences without giving every component the same permissions.
06 / What counts as working
The phone still gets the final vote.
Software checks covered confirmation, corrections, profile boundaries, replay protection and the board seeing the same state. Browser checks covered small screens, long items, cancellation and library navigation. Physical checks established microphone input, spoken output and the phone-to-board request loop.
Recovery also exposed real problems: USB permissions after startup and weather waiting too long to retry. Those were corrected. I subsequently confirmed the tested unplug/replug, phone restart and computer restart sequence.
The newest reader still needs hands-on observation of independent child navigation, narration timing and real speech recognition. Fake microphones and silent test audio are useful for software behavior. They cannot establish what a child can actually hear, understand or operate.
The lesson I keep coming back to is that the useful work often happens between the impressive parts: showing what was heard, letting a child cancel, explaining a failure and preserving a save across a retry. Berry needs those details as much as it needs a good face.