A pit crew that rides along.
Every job runs the same way today: her app on your phone, and a Shiro computer you own doing the thinking — reached through an encrypted tunnel on the road, or straight across your own home network. For the mechanic, a small plug-in reader in the car's diagnostic port feeds your phone. The five steps below are the mechanic's, from the moment you turn the key to the moment she has something to tell you.
On this page: One door · The models · Earned trust · What runs where · Your data · Read-only · Guests · How she compares ›
to the car
its normal
early
plainly
in the cabin
One door for everything you say.
Every tab's question box and mic go to one front door. A table of commands in code is checked first. What it can't place goes to a small model on your computer that can only pick from a fixed list of actions.
- Commands change the app — the station, the route, the cabin (simulated for now) — never the car.
- When she confirms a command, the words come from code describing what actually changed, never from a model.
More detail
Asking about something never changes it: "is the radio on?" gets an answer, not an action. Your phone turns speech into text first; the door works on the words. A question goes to the one part of her that owns the answer, and that part answers on your computer.
Before she speaks one of her own replies, code screens it for numbers it can't find in her data. Replies outside the radio are also screened for lines claiming she did something. A reply that fails is swapped for a line built from her data. If no model is ready in time, she tells you so instead of guessing.
After sorting, a line she can't place may go to one metered hosted AI call. It's visible and switchable.
The models on your computer.
The language models run with llama.cpp, compressed with 4-bit quantization, on your computer's graphics card when they fit. Numbers are computed in code; the models phrase, sort and explain.
| Model | Her job |
|---|---|
| Qwen3-4B-Instruct-2507, fine-tuned by Shiro AI | Her words and personality on every tab, and the first choice for sorting what you say. |
| Gemma 4 E4B-it | Judges every line a guest agent sends; writes the short weather, route and radio summaries. |
| Llama-3.1-8B-Instruct | Keeps weather, route and radio terms accurate; writes the radio DJ's short lines. |
| Phi-4-mini-instruct | Works through the numbers our code computed for weather, routes and radio, and says which mattered. |
| Whisper (base), run with faster-whisper | Turns radio news, and other stations' DJ talk, into text. |
More detail
They load in stages, as her jobs need them. If her fine-tuned model isn't loaded, another model on your computer steps in. Radio news is transcribed on the computer's processor, so it never pushes a language model off the graphics card.
The car's drift detector, weather's own forecasters and nav's arrival-time correction aren't language models. They're statistical. None of them is trusted with what you're told until it has a record. Each is graded against a simple baseline or held in review (next section).
Her own predictions earn trust before she relies on them.
- Nav. Her arrival-time correction is used only on a route, at an hour, where it has proven closer than the map's estimate. Elsewhere, you get the map's estimate.
- Weather. Her own forecast is mixed in only where her graded record beats the published forecast. Today that's nowhere, so you see the published forecast.
- The mechanic. Her own alerts wait in review today; a trouble code the car itself reports isn't held back.
More detail
Nav checks her correction against the map's estimate on drives she hasn't seen yet. Only your own timed drives count; test and simulated drives can't teach her. Her route choices follow the map's order until one of her styles proves itself on your drives.
Weather runs her own forecasters and grades each one against what actually happened. Within her own forecast, a forecaster gets a say only after it beats "same as yesterday" on days it wasn't trained on. One reference forecaster always stays in as the anchor. Her numbers still appear beside the published forecast, labelled as hers, with their grades. The screen labels whose forecast it's showing — hers too, if the published one can't be fetched. Each forecast is recorded the first time she makes it and can't be revised afterwards, so hindsight can't flatter her grades.
Offline-first: what runs where.
This table lists the main parts that go online, not every request. "Hosted AI" means a text model we reach online; "cloud help" means any optional feature that uses it.
Scroll the table sideways to see every column.
| What | Where it runs | Off, or no signal |
|---|---|---|
| Voice to text | your phone's speech recogniser; on many phones, the phone maker's online service | type instead |
| Understanding you, and her replies on every tab | your computer | at home: unaffected; on the road: needs signal to reach it |
| Lines she can't place | one metered hosted AI call | she says she didn't catch it, or asks which part it's for |
| Optional daily summary | hosted AI | no summary |
| Her speaking voice | a hosted voice service | your device's built-in voice |
| Her animated face on the computer's screens | an online avatar service | a locally animated face |
| Radio DJ lines and radio-tab chat | your computer; hosted AI only if the radio's cloud help is on (off today) | your computer |
| Facts for the DJ | a public online encyclopedia | just the song |
| Naming the song on the radio | an online song-recognition service | no song name |
| Guest AI agents (in testing) | the guest's own service, or hosted AI through our meter | no guests |
| Radio stations, weather and radar, place search and routing | online sources: RadioBrowser, the U.S. National Weather Service, Open-Meteo, RainViewer and OpenStreetMap-based services | they stream in when you have signal and degrade quietly when you don't |
| Her music mode, and songs requested by name | an online music service | radio stations only |
More detail
What each part sends:
- Voice to text: your spoken words, from the phone.
- Lines she can't place: that line and the last few lines of your conversation; the answer is screened and introduced as coming from outside.
- Daily summary: a digest with identifying details stripped by code; if a final scan still finds any, nothing is sent. The summary is labelled as written by a model, and it's never used as a fact in her answers.
- Her speaking voice: the text of each new line; a line she has said before plays from a cache on your computer.
- Her animated face: her voice audio as she speaks.
- Radio DJ lines and radio-tab chat, when the radio's cloud help is on: the song, the artist, a short fact and, for songs you ask for and for radio-tab chat, the last few lines of your conversation.
- Facts for the DJ: the artist and song name.
- Naming the song: a few seconds of the station's stream — never your microphone.
- Guest AI agents: the conversation with that guest.
- Stations, weather, search and routing: station searches, the forecast location, places you search for, and a route's start and destination.
- Music: the songs she looks up, and your song searches.
Typing on the road also needs signal: it reaches your computer the same way. Today every call to hosted AI goes through one meter with a spending budget and a daily cap. One switch in the app's Settings turns hosted AI off; her hosted voice has its own switch.
Where your data lives.
Drives go from your phone straight to your computer. Pairing is a one-time QR scan or short typed code that works once and expires within minutes.
- Each drive is checksummed on the phone and checked again on arrival; a file that changed on the way is refused.
- A drive stays queued on your phone until your computer confirms it; a failed upload gets a Retry button.
- The phone app has no ad or analytics libraries of its own; its QR scanner is Google's Play-services scanner.
More detail
The app's drive uploads and live car data are signed with your phone's own key. The computer signs its answers back, so the app knows it's your computer. The in-app screens carry a per-device pass that only a paired device holds.
The tunnel you use on the road is built into the app. The phone makes its own tunnel key, and the app never sends the private half to your computer. Only traffic for your computer goes through the tunnel; the rest of your phone's internet is untouched.
Read-only, and labelled.
The app's link to your car is a fixed list of standard read requests. It has no command to clear a code, reset a module or change a setting, and nothing on your computer can send one.
- Simulated, replayed and stale data each carry their own label on screen. When the mechanic's numbers come from the simulator, she tells you first.
- Every recorded drive is decoded twice, on the phone and again on your computer. On our real test car's sessions: 1,140,958 checks, zero disagreements (counted September 22, 2026).
More detail
Ask her to clear the codes or switch off a warning light and nothing happens to the car. Approving or declining an alert only tells her what you care about.
Every drive is labelled by the wire it came over: your real car or our test bench. A feed counts as your car only when it came over the car's own wire from the car you paired; if she can't verify that, she says so. Once your own car has connected, the demo simulator stops playing scripted faults; what's left is a labelled placeholder with no alerts. A bench recording that claims to be the car is refused.
When she tells you what's usual for your car, that figure comes only from readings verified as your car's own. It never comes from the simulator or a test bench. Until there are enough of them, she doesn't quote one at all.
If a trouble-code read comes back incomplete, the app says "partial read", not "none". A code it missed is never reported as cleared. Her overall health number says how many systems it covers and names the ones it doesn't. A system with no reading is never counted as healthy. Of the five systems in the design, the transmission is measured but has no fault model yet. Brakes, steering and suspension aren't in a car's standard diagnostic data.
The second decoding works from the raw bytes the car sent. Each one-second reading carries many gauge values, and any disagreement between the two decodings is counted and reported. The car's engine computer runs its own built-in tests against the manufacturer's pass/fail limits. In the same count, we read those results in 40 of our real test car's sessions: zero failures. An answer we can't decode with certainty is set aside, never counted as a pass.
Guests follow house rules.
Guests are outside AI agents invited to talk with you on her Deck, a supervised spot on your phone. A guest agent has to accept Shiro's written house rules before it says a word. Every line it sends is checked twice: fixed rules in code, then a judge model on your computer.
- If the judge can't answer, the line is held. Nothing reaches you unchecked unless you let it through.
- A guest gets no buttons and no tools: text only.
- Trying to work the car breaks a hard rule: the guest is disconnected, flagged and its pass revoked, and the conversation is set aside.
More detail
Only a model that passed our calibration test may judge. A softer slip — too long, flooding, tone — gets one warning, and the line isn't read to you (it stays in the log, marked withheld). A second slip ends the visit. She reads their lines out loud to you in a separate guest voice (you can turn it off) — never her own.
She can also call an AI agent you've saved yourself (in testing): a link to the agent's own service, not a phone call. While you're on the call, car commands, questions about your trip and your saved places stay with her. Everything else is privacy-scanned before it's passed on. On a call there's no letting a line through by hand: if the judge stays unavailable, the call ends.
Ending a visit is one tap, and on a call, "hang up" always gets through.
Next to what's already out there.
Categories, not brands; every cell is sourced (below).
Scroll the table sideways to see every column.
| Shiro | Code readers & OBD apps | Phone & cloud voice assistants | Automaker connected services | |
|---|---|---|---|---|
| Reads your car's own data | yes: engine, emissions and electrical, through a plug-in reader TODAY | yes: codes; live data on many | mostly no; some in-car versions show basic status | yes, for that maker's cars |
| Flags drift before a warning light | built to; shown so far on simulated faults SIMULATED | mostly once a code is pending or stored; varies | no | varies; fault alerts and service reminders, a few predict parts |
| Adapts to your specific car | gathering our real test car's readings now BUILT-TO | mostly no; a few apps track trends | no | varies |
| Explains readings for drivers, DIYers and mechanics | yes, three views; the site's samples are simulated TODAY | code definitions; some apps explain and suggest fixes | general answers; some quote the owner's manual | app alerts and reports; varies |
| Works with no signal | partly: drives keep recording; away from home, her answers, radio, weather and maps need signal TODAY | reading yes; app extras often need data | mostly no; a few basic requests work | dash warnings yes; connected features need signal |
| Your driving history stays on hardware you own | yes: your phone and computer; cloud help sends only what it needs TODAY | handhelds yes; apps often sync to their maker's cloud | mostly the provider's cloud; some on the phone | collected by the automaker; sharing varies |
| Can change the car or clear codes | no: read-only by design TODAY | most can clear codes; some change settings | some send climate or remote-start commands; varies | remote commands; some push software updates |
| Account or subscription needed | not today; hosted AI is optional and metered TODAY | one-time buy; apps may need an account or subscription | usually an account; some paid tiers | account; many features paid after a trial |
| Car health in the same assistant as radio, weather, maps | yes, one app; the cabin tab is simulated TODAY | car only | music, weather, maps; not car health | varies |
Only Shiro has all four of these together today: your car's data, read-only, on hardware you own, in one app with radio, weather and maps.
Sources
Comparison sources: SAE J1979 (OBD-II diagnostic services); U.S. EPA and the California Air Resources Board (OBD rules); NHTSA (vehicle cybersecurity); FTC (connected-car data); Mozilla Foundation's *Privacy Not Included car review; and each category's own product documentation. Product names withheld.