How wingmin decides
A plain account of what runs where. Updated 8 October 2026.
wingmin runs on your devices (iPhone, Apple Watch, Mac, Windows PC), in a web app you host, and on a small server we run. This page says what each one does, what each speech engine sends where, and how AI requests are guarded.
On your devices
- Recording. Always started by you. The Watch starts from the app or the Action Button and shows a haptic and an indicator; a phone call or Siri interrupts it, and the Watch says so. On Mac and Windows you choose in the room (microphone) or call (microphone plus the call’s sound, kept as two channels, “me” and “them”). When a call app starts using the microphone, the Mac or PC offers to listen; it never records without a click.
- The Watch on its own. The Watch has no speech recognition of its own (Apple doesn’t ship it for watchOS). With the iPhone in reach, audio goes to the phone. Without it, the Watch uploads short audio pieces over Wi-Fi or cellular to a speech service, or records locally and hands the recording to the iPhone later. You can pin “never use cellular”.
- Live matching. Flags, spoken capture (“note that…”, “remind me to…”), keyword watch and speaker-introduction suggestions are matched on your device, from the words, with no AI call.
- Storage. Conversations are stored on the device and synced between Apple devices through your own iCloud private database. Notes, ask answers and generated pages sync through your vault when one is connected.
- Secrets. Detection of passwords, card, account and ID numbers and keys runs on the device, with pattern rules plus Apple’s on-device model on iPhone and Mac. It never runs on cloud AI.
On wingmin’s server
The server holds accounts and usage counts and relays requests. It keeps no conversations, transcripts, notes or vault passwords.
- Sign in with Apple gives a stable, pseudonymous account id for the monthly usage limit and the list of your signed-in devices.
- AI relay. The request (already redacted on your device) is forwarded to a model through our router and the answer returned. The text is held in memory for that request only. As a second guard, the server refuses any AI request that still contains a secret-looking pattern or a note marked private.
- Speech. The server issues short-lived keys so your device can stream to the speech service directly, or relays the audio when that isn’t possible. Relayed audio is forwarded, not stored.
- Vault handoff. When a Windows PC asks your iPhone for the vault connection, the iPhone seals it for the PC’s one-time key; the server holds only that ciphertext, in memory, for at most five minutes, and hands it over once. Both screens show a short code computed from the key, so a swapped key would show a different code.
If the server is switched off, the apps still work with your own keys or on-device options.
What each speech engine sends where
| engine | where the audio goes | speaker labels | notes |
|---|---|---|---|
| Apple (on device) | Stays on the iPhone or Mac | No | Needs the iPhone present for Watch recordings. |
| Whisper (on device) | Stays on the iPhone or Mac | No | Used for mixed Mandarin and English. One-time model download. |
| Deepgram (default) | Deepgram, directly with a short-lived key or your own key, or through our relay | Yes | Every request asks Deepgram not to use the audio for training. |
| AssemblyAI | AssemblyAI, directly with your key or through our relay | Yes | Chosen by hand under advanced. |
| Your own speech | The OpenAI-compatible service whose key you entered | No | On a call, the two channels still separate you from the other side. |
| Free on this PC (Windows) | Stays on the PC | After the meeting | Models downloaded on first use. Early build. |
Before audio goes to a cloud engine for the first time, the app asks and names the provider. Comparison mode sends the same audio to two engines at once and names both.
How AI requests are run
- Three routes, in order: wingmin’s server, your own AI provider, then Apple’s on-device models. Results depend on the model.
- One gate. Every request to cloud AI, through our server or straight to your own provider, passes through one place in the app that redacts secrets and drops notes kept off cloud AI. On-device AI gets your real text.
- Transcript text is data, never instructions, in every prompt.
- Grounded. Key points, decisions and tasks link to the transcript moment; ask wingmin answers cite the conversations and notes they came from.
- One switch in Settings turns all AI off. Recording, transcription and flags keep working.
What a flag is
A flag is a timestamp in the transcript with a type (important, follow-up, question, decision) and the words around it. You place one by hand, or wingmin places one when the live transcript matches a phrase from your list (“what do you mean by that?”, “so we’re agreed”, “I’ll send it by Friday”), a keyword you watch, or a spoken capture. The analysis reads the transcript and the flags, and treats the flagged parts as worth attention.