Norbert Klockiewicz

On-device AI and DevRel @ swmansion | React Native ExecuTorch | Private Mind | Velonos

[Norbert Klockiewicz @Nklockiewicz]

I built a 3D scanner in React Native that runs fully on-device :D

You take 8 photos of an object or scene, Depth Anything 3 turns them into 1.2 million points you can spin with your finger, ~3s on an iPhone 16 Pro. The model even figures out where the camera was standing for each photo.

Watching @theworldlabs drop their Atlas world model two days ago sent me down this path. That's spatial intelligence at cloud scale and I wanted to see how much of it already fits on your phone.

stack: @swmansion's react-native-executorch + typegpu

[Norbert Klockiewicz @Nklockiewicz]

@jullerino The ability to choose if the dictation runs on device or in cloud is super cool

[Software Mansion @swmansion]

When you ask Claude to summarize an email thread, every name, address and IBAN goes with it.

Velonos redacts them on your device, so the AI only sees placeholders, and the answer comes back with the real details filled in.

🔒 https://t.co/ZlkBtOuZDc

[Norbert Klockiewicz @Nklockiewicz]

@peijiancc Thanks for sharing Private Mind! Glad that you liked it. If anything breaks, GitHub issues are the fastest way to reach us.

[Norbert Klockiewicz @Nklockiewicz]

@notaryannnn @Qualcomm Model inference only, the whole pipe adds around 10-15 ms.

[Norbert Klockiewicz @Nklockiewicz]

I've run the same depth model on the NPU and the CPU of a Galaxy S25: 27 ms per frame vs 550 ms :D

If the phone has an NPU, use it. Kudos to @Qualcomm for this chip, the 8 Elite even beats the iPhone 16 Pro's Neural Engine (~90 ms). In the next react-native-executorch release most of our models run on the NPU on Snapdragon devices. Android isn't a second-class citizen anymore.

[Norbert Klockiewicz @Nklockiewicz]

@Qualcomm You can see how much of a latency there is in xnnpack version, completely unusable and it throttles faster then NPUs.

And of course the visuals are inspired by work of one and only @reczko_konrad.

[Norbert Klockiewicz @Nklockiewicz]

@G_Moskaliuk @tomekzaw_ Combining the potential of talented programmers with the capabilities of agents yields incredible results; it’s fascinating to wonder how much more efficiency might be squeezed out of this.

[Norbert Klockiewicz @Nklockiewicz]

@ReactNativeRwd Thanks for the 0.11 shoutout! Working with on-device LLMs in react-native-executorch should be much more stable now.

[Norbert Klockiewicz @Nklockiewicz]

@Alanogrok Agreed. We are working on benchmark that will be published in near future. It will give you a look into both recall and precision of our system as well as some other interesting statistics.

[Norbert Klockiewicz @Nklockiewicz]

Finally I can share with you a thing I was working on for a past few months.

It's called Velonos and in a few words, it's a privacy gateway that filters the traffic that goes to chatGPT, Claude and Gemini, scans the documents and prompts and hides a sensitive information and restores it in the answer when it comes back.

Most important part is that everything happens on the device.

Here's a link and I will share some more details in thread.

https://t.co/6htN0mXbXR

[Norbert Klockiewicz @Nklockiewicz]

@HeathGerrald Based on context. A fine-tuned model reads the whole sentence, so “Jordan signed” gets hidden and “shipping to Jordan” doesn’t. Brands and products stay visible on purpose. False positives get scored like leaks.

[Norbert Klockiewicz @Nklockiewicz]

To detect sensitive information we are using a model that was fine-tuned by us and combined it with a static rules. This hybrid approach gave us one of the best systems for pii detection that can work on a laptop with 8 GBs of RAM.

[Norbert Klockiewicz @Nklockiewicz]

In the response, LLM will use those placeholders, so it will answer with something like this:

"Hello [PERSON_1], nice to meet you!"

Before this message is shown to user, Velonos restores the original values, so in the ChatGPT user will see:

"Hello Norbert Klockiewicz, nice to meet you!"

[Błażej Kustra @blazejkustra_]

Yesterday we announced Argent e2e

Today: how to get it running on your CI, start to finish in under a minute 🎬

[Norbert Klockiewicz @Nklockiewicz]

I mean… WHAT?!?

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Kamil Kaczmarczyk @_kkaczmar_]

Searching a 54-min podcast by meaning, in any language, fully local in the browser.

EmbeddingGemma 2 matches your question straight to the audio, no transcription. Ask in Italian or Chinese, it still finds the right moment.

Runs on your GPU with ruNNtime (WebGPU)

[Norbert Klockiewicz @Nklockiewicz]

@_kkaczmar_ 🔥🔥

[Norbert Klockiewicz @Nklockiewicz]

@_JoshuaTimothy @Snapdragon @Qualcomm_Dev @liquidai @Qualcomm Sick! Do you maybe know how much faster is this model when it runs on qualcomm NPU instead of CPU(xnnpack for example)?

[Norbert Klockiewicz @Nklockiewicz]

@austingriffith That's a great use case for on-device AI, did a similar thing with local VLM and foreign menus. Just don't forget to download the model before you fly ;)

btw. what are you running the models with?