Tried this and really liking it so far. Question - is there a diarization support in the tui app or any of the models MetalRt supports? Any plans to add it if not already supported?

Yes, we do have plans to support it.

Seems to be based on See through walls by MIT (2013)? Good job porting it to esp32. I was just looking a week ago to do the same thing - basically reproducing this work. I'll definitely try this. https://www.media.mit.edu/projects/seeing-through-walls-comp...

"ESP32 CSI Toolkit", papers+code, Virginia Commonwealth University, USA https://stevenmhernandez.github.io/ESP32-CSI-Tool/

"Wi-ESP", papers+code, KITS WRL, South Korea, https://wrlab.github.io/Wi-ESP/

"Espressif CSI sample applications", Shanghai, China, https://github.com/espressif/esp-csi

"CSI-MURDER", paper+code, WiFi Sensing countermeasures, EU, https://ans.unibs.it/projects/csi-murder/

I believe that was the initial paper which really started the Wi-Fi sensing research. Although a lot of research has come after that. It's a really fascinating technology with a whole range of possibilities beyond just motion and presence sensing.

I was reading Xiaomi YU7 marketing page[0] yesterday and the NVIDIA AGX Thor stood out (says: NVIDIA DRIVE AGX Thor). I was wondering what it was and this showed up! Looks like it is (or a Drive variant of it) is already being used in newer cars for self-drive and such. [0] https://www.mi.com/global/discover/article?id=5174

Can someone (or OP) point me to a recipe to fine tune a model like this for natural language tasks like complicated NER or similar workflows? I tried finetuning Gemma3 270M when it came out last week without any success. A lot of tutorials are geared towards chat applications and role playing but I feel this model could be great for usecases like mine where I am trying to extract clean up and extract data from PDFs with entity identification and such.

If you're really just doing traditional NER (identifying non-overlapping spans of tokens which refer to named entities) then you're probably better off using encoder-only (e.g. https://huggingface.co/dslim/bert-large-NER) or encoder-decoder (e.g. https://huggingface.co/dbmdz/t5-base-conll03-english) models. These models aren't making headlines anymore because they're not decoder-only, but for established NLP tasks like this which don't involve generation, I think there's still a place for them, and I'd assume that at equal parameter counts they quite significantly outperform decoder-only models at NER, depending on the nature of the dataset.

This is using the gemma-llm python library which uses JAX in the background: https://gemma-llm.readthedocs.io/en/latest/colab_finetuning....

Have you tried this one here by any chance?

https://huggingface.co/dslim/bert-base-NER

Just wondering if it’s worth testing and what it would be most useful for.

FYI: owhisper pull whisper-cpp-large-turbo-q8 Failed to download model.ggml: Other error: Server does not support range requests. Got status: 200 OK

But the base-q8 works (and works quite well!). The TUI is really nice. Speaker diarization would make it almost perfect for me. Thanks for building this.

we store data in R2 and range query sometime glitch... It might work if you retry it

I tried a picture with instructions and it says "something went wrong". I would love to try and see how well it works for my use case.

I'm so sorry you ran into that, and thank you for reporting it. This is exactly the kind of feedback I need at this early stage.

You're right, my backend logs show that most requests are succeeding, which means there must be an error happening somewhere between the front-end and the server that I'm not catching properly yet.

Based on this, implementing a more robust error logging system is now my top priority. I'll get on it right away so I can find and fix these issues for everyone. Thanks again for giving it a try.

Can you open up the options to use other model/versions, especially Gemini-2.5 pro experimental models available through aistudio? Would love to try this but gemini flash fails for even simple tasks. Example: I asked it to extract all the links from comment section of a hackernews comment section and it just scrolled all the way to the end and then nothing. Maybe pro models can do it better.

"Gemini Flash fails even for simple tasks." On the Gemini Flash page (https://deepmind.google/technologies/gemini/flash/), it claims to be 'best for fast performance on complex tasks.'. I always use Gemini Flash in my project for demos and testing, and it performs very well, if a project requires a large, expensive model to handle simple tasks, that could be an issue to users.

Gemini 2.5 pro is available. Is it missing on your side? Do you run index via CLI?

Yes it is, however API keys from aistudio only allows pro-experimental model. So if I select gemini-pro, I will see this: "Gemini 2.5 Pro Preview doesn't have a free quota tier. Please use Gemini 2.5 Pro Experimental (models/gemini-2.5-pro-exp-03-25) instead". Can I choose exact model somewhere in the CLI?

Oh I see, didn't know about that, fastest and easiest thing you can do is to play around with pro via our chat UI https://lmnr.ai/chat - it's free up to 10 messages.

For the CLI and custom models, you can clone the repo, then go to the cli.py and manually add your model there. I will work on proper support of custom models.

extremely slow

which model are you using? try gemini pro/flash, they are very fast

This is really cool (and timely for me). Lovely work with the UX. No accounts, no nonsense. Kudos.

> Spends so much time looking for UI design tools that they could've learned actual design in the meantime

Ouch! This is really well done. :)

I am RM2 user and this is useful for me. I signed up not expecting much but that digest designer is so well done! Kudos. Beautifully executed. Signed up and looking forward to the digests.