-
-
What wali remembers: lasting facts about you, built up from your conversations over time
-
wali pulls out the promises you made today, like a 9:30 rehearsal, and flags when it's not sure who said it
-
Approval cards: wali asks before acting for you, and follows up on plans it heard
-
iPhone app: pendant status, last conversation, and a quick way to find your pendant
-
The wali web app: chat with an assistant that remembers what it's heard
-
Chatting with wali on the web, grounded in your real conversations
-
Connect your apps through Composio so wali can take action, with your OK
-
Meet the pendant: a small wearable you just put on and forget about
-
A Fitbit for your mind: the wali landing page
Inspiration
Most of what we mean to do gets said out loud and then forgotten: "I should call her back", "remind me to send that", "I keep losing hours to my phone". We wanted something that hears those moments and quietly helps, without being one more app you have to open. It also had to feel calm and human, not like surveillance.
What it does
wali is a small pendant you wear on a cord. It picks up what you hear and say through the day, and your phone sends the audio up to be transcribed. After each conversation wali decides whether it's worth speaking up, and most of the time the answer is no. When it does speak up, it's a short, specific nudge on your phone, a follow-up it schedules, or an offer to handle something for you. Anything that reaches people, spends money or deletes things waits for your OK on an approval card.
In the iPhone app and the web app you can:
- chat with wali, which remembers what you've told it and what it has heard
- read and search your transcripts by day, with playback
- see what wali remembers about you
- connect apps (through Composio) and let wali act in them or in a browser, with your approval
- choose when wali speaks up (quiet hours, how often)
- try it without the pendant by pasting in a conversation
How we built it
- Pendant: ESP32 firmware streams IMA ADPCM audio over Bluetooth LE using our own protocol, and backfills audio recorded while the phone was out of range.
- iPhone app: Expo / React Native with native iOS controls. It saves every audio frame to disk first, cuts the stream into conversation segments, builds WAV files and uploads them to Convex. It's built to recover from crashes, restarts, going offline and iOS suspending it in the background.
- Transcription: ElevenLabs Scribe, with speaker labels, run from Convex.
- Deciding when to speak: rules come first (quiet hours, caps, a cooldown after you dismiss a nudge). Then a gate model (Jev, through OpenRouter) decides whether anything is worth saying. Then a Mastra agent chooses between nothing, nudge, ask first, act, follow up and support.
- Memory: Supermemory, one profile per user, built from finished conversations and the notes wali saves.
- Actions: Composio for app actions and Browser Use for web tasks, behind permission modes and approval cards.
- Web app: Next.js, shadcn and Clerk sign-in, with push notifications through Expo and Web Push.
Challenges we ran into
- Keeping audio flowing while the phone is locked. iOS freezes timers in the background, kills apps that use too much CPU, and can relaunch an app without any screen. We moved segment closing and uploads onto incoming Bluetooth events, cut idle CPU from about 20% to about 1%, and start the pipeline without a view.
- Not losing audio. An unawaited file write caused missing WAVs, so every write is now synchronous, and broken segments rebuild themselves from the saved frames.
- iOS held uploads started in the background for up to 100 minutes until we switched to a regular session.
- After a Wi-Fi/cellular switch the app could wait on a dead connection for up to a minute, so we wrote our own reconnect handling.
- Knowing when not to speak. A proactive assistant that talks too much is worse than none.
Accomplishments that we're proud of
- A full path from a pendant on your chest to a useful nudge on your phone.
- About 30 integration tests that run the real recorder, uploader and Bluetooth code against a simulated pendant and server.
- iPhone and web apps with the same features on one backend, and a warm, calm design.
What we learned
- Saving first and processing later makes a hardware pipeline far more forgiving.
- iOS background rules shape the whole architecture.
- The hardest product question is when to stay quiet.
What's next for wali
- Tying recordings to each pendant's owner.
- TestFlight builds for more testers.
- Tuning when wali speaks up from real days of use.
Built With
- bluetooth
- browser-use
- clerk
- composio
- convex
- elevenlabs
- esp32
- expo.io
- mastra
- next.js
- openrouter
- react-native
- shadcn
- sqlite
- supermemory
- typescript