Daniel Duke (@iamdanduke) header
Daniel Duke (@iamdanduke) profile picture

Brother, creator, award-winning singer-songwriter. Leading marketing and content at @desertantlabs. Building tiny apps at https://teapawt.org/t.co/ryOa1HnytE.

North of Ireland
Joined April 2008
Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)Photo by Daniel Duke (@iamdanduke)
Very excited to have joined this team.
Today we're launching Desert Ant Labs: a European frontier AI lab building on-device intelligence. 18 models across audio, vision, and text. SDKs for Swift, Kotlin, and JavaScript. No tokens. No logins. Nothing leaves the device. desertant.com
8
346
Daniel Duke retweeted
OpenMates, an open-source AI assistant, checks messages for personal data before they're sent to the AI. It runs our Redact model in the browser to find names and addresses and swap them for placeholders. openmates.org
1
1
7
469
Our language identification model, Ear, covers 102 languages. It named the right one from 5 seconds of speech in each of these Spoken Wikipedia recordings. Run it before you transcribe, so each recording goes to the right speech model on a user’s device. desertant.com/models/ear
2
7
72
3,131
Transcribe a podcast recording at 35,000 feet. ✈️
10 minutes of audio, transcribed in 1.4s, in airplane mode. Voz, our speech-to-text model, ran this on an iPhone 17 Pro with no connection. Every word comes back with a start and end time, which the video uses to highlight each word as it plays. 25 languages, with SDKs for Swift, Kotlin and JavaScript. The recording never leaves the phone, and there’s no fee per minute of audio. desertant.com/models/voz
2
169
Daniel Duke retweeted
Fresh in the MacPip directory 🖥️ Hermes by @iamdanduke: Switch your bluetooth headphones between Macs from the menu bar. $3.99 once.
2
4
131
Build smarter chat, on device. Four models you can add to your messaging app today: transcribe voice messages, suggest emoji, catch bank details before they’re sent, and detect languages from as few as three words. It all runs on the user’s device, with no cost per message. Every result in the clip is real model output. desertant.com/use-cases/chat
1
5
413
Build a whiteboard or notes app with pixel-perfect shape snapping. No cloud calls required. Shapes, our shape recognition model, turns a rough stroke into a clean shape, on device. It now supports 13 shapes, arrows included. Scribble over another shape to erase it. Under 1MB, with SDKs for Swift, Kotlin and JavaScript. Works with tldraw and Excalidraw stroke data. desertant.com/models/shapes
2
1
9
657
If a company is too big to fail, then it’s too big to remain privately owned. You can’t expect taxpayers to carry the risk while shareholders enjoy all the rewards.
60
1,437
22
10,961
96,959
Uhm, our filler-word detection model, now runs on Android and the web. It marks the "um", "uh" and "hmm" in a recording to within 20ms, without needing a transcript. SDK 4.3.0 adds it to our Kotlin and JavaScript SDKs, and the Swift SDK now runs it on Linux and Windows too. On Android and the web it finds the same fillers as on an iPhone or Mac: 635 of 642 in our test on 2.1 hours of podcast audio. desertant.com/models/uhm
2
1
1
16
1,303
Catch abuse and hate speech before it's posted, on the device it was typed on. Toxic scores a reply or message before it gets sent, so your app can warn the writer or hold it for review. On an iPhone 17 Pro that takes 1.8ms, and nothing leaves the device to be checked. One model covers 35 languages. Toxic is in beta in our Swift, Kotlin and JavaScript SDKs. desertant.com/models/toxic
1
1
6
706
Daniel Duke retweeted
I read somewhere that we’ve moved from designing by stacking blocks, almost like building with LEGO: piece by piece, adding elements until a form emerges. With AI instead, we can quickly begin with a block of possibilities that we then have to polish, carve away, and refine. There’s something similar in Florentine Neoplatonism, which held that the idea exists before the material. The sculpture already exists within the block of marble; the artist’s job is to remove what conceals it. The risk, perhaps, is stopping halfway.
6
2
22
2,882
One of my favourite things about building a Mac app is having an excuse to come up with a clever app icon. 🔊 + 🪽
Hermes is now on the Mac App Store! 🎉 I built a little menu bar utility to switch my cheap Bluetooth headphones between my work and personal Macs without having to open Control Centre. If you also have this first world problem, Hermes is for you. apps.apple.com/us/app/hermes…
4
5
371
This is a such a great example of what's possible with @desertantlabs models. A whole meeting transcription app running on 8 tiny models! 🐜
Murmur, by @uwe_, is a free Mac app that transcribes your meetings and interviews, all on device. Murmur runs eight of our models one after another: Ear detects the language, Voz transcribes, Who labels the speakers, Align times each word, Tongue checks the language of the text, Uhm marks the ums, Clear cleans up the audio and Redact hides names and numbers. Uwe's own tests put ten minutes of audio through the pipeline in about ten seconds on an M2 Mac. uwe.social/murmur/
1
133
Daniel Duke retweeted
For years I waited for Siri to actually do things. Eventually I stopped waiting and built the first assistant that lives on your iPhone. Pelican runs routines, works while your phone is locked, and uses the apps you already have. No Mac mini necessary. Out today. ⏰ Routines run on a schedule, even with your phone locked in your pocket. A 7am brief with your calendar, the weather, and anything urgent in your inbox. Or a price watch that only pings you when it drops. 📱 Pelican works in your existing apps: adds events and reminders, turns off the lights, sets alarms, and drafts texts and emails for you to send. With Shortcuts support it can run any app intent from any app on your phone. 📸 It finds any photo by what's in it. Your photo library is analyzed by an on-device model and never leaves your device. 🧘 Ask it something big and lock your phone. Pelican keeps working on your lock screen with a live activity, and the reply arrives as a notification. 🍎 Use Apple's model (free, no account), Claude, ChatGPT, Gemini, or a self-hosted open-weights model running on your own infrastructure. Your chats never touch my server. Free on the App Store 👇
17
8
3
151
21,329
Hermes is now on the Mac App Store! 🎉 I built a little menu bar utility to switch my cheap Bluetooth headphones between my work and personal Macs without having to open Control Centre. If you also have this first world problem, Hermes is for you. apps.apple.com/us/app/hermes…
6
1
1
11
1,153
If your iPhone grabs your headphones first, there’s a fix for that too. Hand Off is a free iPhone Shortcut that turns your phone’s Bluetooth off for five seconds, so your Mac can steal them. Pro tip: put it on your Action Button. 🪽 Set it up here: littlevictori.es/hermes/help…
79
Daniel Duke retweeted
HE OFFERED THE WRONG DANA WHITE THE JOB ON THE SIX YEAR ANNIVERSARY OF FOUR SEASONS TOTAL LANDSCAPING SOMEONE PINCH ME
124
3,845
75
64,570
748,512
AI emoji suggestions you don’t have to wait for. Emo suggests emoji based on what you type, on every keystroke, on device. On an iPhone 17 Pro, Emo takes 1.7 ms per suggestion, fast enough to keep up with you. 22 languages. 5MB. Nothing leaves the phone. desertant.com/models/emo
1
1
10
544
Typing to your agent is so 2025. Here are three dictation apps built with Voz, so you can talk instead. ∗ Takt puts push-to-talk on Windows: hold Right Alt, talk, and the text lands in whatever app you’re in. Clear cleans up the mic audio first. ∗ Whisper Anywhere does voice typing from the Windows tray, with Voz as the default, and never writes a recording to disk. ∗ OpenFlow brings it to Android as a floating bubble that shows up with the keyboard, with Ear detecting the language. Every word is transcribed on device. Links in the thread.
2
1
18
894
Half as many false flags as NudeNet. Moderator is a 9.7MB NSFW detection model that can run on a phone before a photo reaches your servers. In our tests, it wrongly flagged 6.3% of safe images, against NudeNet’s 12.9%, and caught 87.8% of NSFW images. desertant.com/models/moderat…
1
3
1
37
1,766
Astra vs Opus 5.5 is such a stark contrast. Astra feels like directing an intern, Opus 5.5 feels like collaborating with a peer. Astra can't even do consistent work in an existing codebase and it just devours usage limits.
1
1
160