Sergey Khalil

Does anyone know if there is a way to go around the WKWebView cold start delays for a webapp wrapped in an iOS native shell?
First start is painfully slow, and I don't really see any framework being able to optimise it beyond pre-warming or WKProcessPool

POV: your todo list is 1 year old 🤨 🧐 🤔????

±4% of iPhone 14's CPU is how much it takes to run state of the art live transcription!

ps: Transcripto will be out soon, I'm back

StackOverflow got eaten, end of an era

I'm working with transcripts a lot lately, @jedutchy asked me if we could improve podcast listening experience for language learners...

I said: "Check this out 😊:

What are your best Siri hacks?
Or other voice assistants

I’ll start: asking Siri to set timers like “<activate Siri> timer 10 minutes” and adding calendar appointments “dentist at 3pm on <this date>”
These save me hundreds of taps per week for sure!

I disable notifications by default on all apps because they're being abused for ads and spam with upgrade requests.

I take users' trust to allow notifications seriously and my apps will never send spam/ads, only genuinly useful updates!

I want the first version to be out before Christmas, so I'm compromising a little bit for now, but I bet this edge case will be quite rare

State of polish: Transcripto will notify you when transcription is done, and warn you if there is a risk iOS will stop the transcription in the background (for very long files) 😎

ps, there is a way to let the transcription run longer, in the background but

Some companies leave messages in browser developer tools, like:
- Threads: warning about self-xxs (you can lose auth cookies by pasting malicious js code into console!)
- Airbnb & Medium: hiring ads 😄

I recall there were more companies that did this, but can't find rn 🤔

A few posts of mine got some attention on Threads, so my wife @jedutchy got competitive, started posting a lot more, and actually managed got double my views in the last 2 weeks 👏🙌🥳

Checked out their license, seems permissive enough, but they can change it any time, which I don't think is standard on OS.
Anyhow, here's a summary table from Claude Opus 4.5 and a nice slide with it from Nano Banana Pro (a bit too corporate for my taste, but not too shabby!)

Introducing SAM Audio: The First Unified Multimodal Model for Audio Separation SAM Audio transforms audio processing by making it easy to isolate any sound from complex audio mixtures using natural, multimodal prompts — whether through text, visual cues, or marking time segments.

From time to time Meta releases something cool, I think their SegmentAnythingModel (SAM) is a great contribution to the OS community: ai.meta.com/blog/sam-au...

What's up with google having soo many ways to interact with their models? (gemini, Flow, aistudio, notebookLM... other?)

Did you know that you're missing out on 75% of iOS users if you only support iOS 26?

By supporting iOS 18, you can 10x (!) your potential users, and if you go lower you can cover pretty much all devices.

ps: iosref(.)com has cool iOS stats, shoutout to the author @eugenemsp

Apps are like 5% core functionality and 95% polish ✌️

Actually, here it is, Transcripto can now transcribe audio messages shared from WhatsApp, Telegram and Signal 😎
(at the very least!)
(sorry, my phone is in Dutch for learning purposes haha)

Optimised Transcripto's launch time to the point that "Loading models" now looks like a bug 😄

Will be removing the model loading screen after the first launch (during the first launch Apple Neural Engine needs to compile the weights for that particular device, takes ±20s)

Left my laptop for a sec while Claude Code was churning, and came back to this 👀 😊

Coreml model storage format is crazy efficient, this bundles a full Parakeet v3 0.6B model.
If you wanna run it locally with python's apple silicone framework the model alone is 2.5GB

They built a VR contraption for rats and it actually worked! Wow
ratsplaydoom.com/

I really admire practical engineering, so simple and to the point!
they even trained rats to shoot, which is slightly concerning haha

Transcripto's live captions for 🇳🇱 Dutch are nearly flawless, and pretty much realtime, Dutch learners, you'll soon have quality subtitles for any audio on your laptop or phone!

it also works for 25 languages other languages (on iPhones too)
mac/ios app is coming soon ✌️😄

Transcription is working at ~100x ⚡️ realtime speed both on Macbook Pro M2 as well as iPhone 14 ✌️

Still ironing out the kinks in token streaming / audio chunking, but it's really coming together! I got suuper granular timestamps too - per token, super cool tech

Seems like the only way to ensure ultimate performance and UX is to build things internally: parakeet 0.6B coreml model loads in 160ms ⚡️
After ANE cache, of course

Goal -> build the fastest/most accurate _realtime/live_ transcription app out there ✌️

iOS devs, I got a question: what’s the fastest/most reliable way to get realtime transcriptions on iOS 17+?
I’ve gotten mixed success with FluidAudio (parakeet), doesn’t seem like the accuracy is high enough.
iOS 26’s transcription is pretty good, but older devices are out…

Since when is it a crime to sign up with email on your own domain name? 😕
KlingAI giving this warning, ensuring pretty bad UX for power users, seemingly in the name of preventing abuse of their free tier/sign up offer?
It's not complicated to get gmail/yahoo emails these days ><

Just found another Privacy Policy page for Gemini products, seems like there are some controls, but oh my, I can't imagine anyone actually being able to stay on top of these

Asked Claude Opus 4.5 to analyse all of Gemini's privacy policies wrt training their AI on the code that you provide as a user - reasonably sure they can do it regardless of the settings 🤷

I start to think that Cursor sometimes forces devs to update with interface glitches

or maybe it's just me having 29 projects open in parallel... ? ✌️ 😅

From first commit to published app in 11 days 😌✌️