properbrew

Love seeing people building hardware. After reading some of the case studies, using nanobubbles seems like a no brainer!

Well this looks like some good stupid fun, I want it. Hope you make an Android version at some point.

Also great video, it hooked me.

I fell down the rabbit hole of voice transcription about a year ago, always had a love for utilising fine tuned LLMs so have put two and two together and built https://whistle-enterprise.com. The biggest challenge being it all running on CPU with the target device being your low to mid spec office laptop that's a few years old (I5, 8gb RAM). All nicely packaged together in a single completely offline selfcontained app that you just install and run (no environment setups, packages to download, models to download etc).

One of the hardest parts I've found is the diarisation (who said what) side of things. Trying to tune this and have it working in a way that doesn't absolutely grind the laptop to a halt or take forever to complete has been _hard_ but also extremely rewarding.

Another part has been the fine tuning side of the Phi-4 model, I'm on version 10 now, getting that pipeline down was a journey in itself, but I've got some great results. I wrote a bit about it in a comment here - https://news.ycombinator.com/item?id=48385906#48389625

I absolutely love working on this, I still wake up and the first thing I think about is voice transcription pipelines (sad I know), but I'm excited to see how much further performance and utility I can squeeze out.

I hate to do the "you're holding it wrong" trope, but I think you might have something misconfigured somewhere unless you missed a 0, because just past 60k tokens is such a small context window to be seeing issue in.

Do you have any old documentation that it's picking up and referencing? If you set all claude settings back to default do you see the same issue?

Thank you! No they cannot tell. It is your requirement as per the laws of your country to notify the other party if you're going to use it.

Thank you for the assumption, I'm actually not a developer at all.

I'm from a hardware / networking / infrastructure background. I've had extensive exposure to (web) application development as I'm working closely with development teams and I do have the bash/powershell scripting knowledge.

But honestly, if I tried this "the old fashioned way" it probably would have taken me about 6 to 7 years to develop that application, that's an optimistic estimate. You really do have to have a passion for what you're building, I didn't know that voice transcription and local LLMs would be such a driving force for me, but it's all I think about, so much that I find it hard to go to sleep sometimes.

I used LLMs to develop Whistle Enterprise (https://whistle-enterprise.com) from the ground up, from scratch.

It's taken _a lot_ of time and effort, but this is an example of what can be developed using LLMs alone.

You have to have dedication and a goal to reach, but you can absolutely build anything if you're building with the right foundations in mind.

Well yes, but if I wanted a completely flat lifeless page I would just use excel and turn off borders.

This one works well. I think it's because there's no shine to it, it's just the data, what you need, right there without trying to fluff it all out with rounded edges and superfluous stuff.

I find it such a hard thing to quantify, I know it's not helpful but you can just feel the slop seep through.

I'm not sure if it's because I've iterated through so many sites that LLMs have produced that "slop" is instantly recognisable and it just feels soulless.

Not like web pages ever had a soul, but it's not there on the generic LLM generated sites.

Ahhh ok I totally understand what you mean. Yea the edge cases are absolutely where you start to feel the pain and things look good on the surface until you dig in. I think even in the age of LLMs the adage of 90% of the time is spent of the last 10% will ring true.

Sure an app can be built and spun up in an afternoon, but are you willing to spend another 6 months ironing out all those little bugs, tuning it a bit, testing, tweaking, testing etc.

This absolutely fascinates me. I had a friend who needed subtitle files generating for audio and using in CapCut yesterday yet none of the available stuff was suitable, so he asked if I could adapt some of my software to export subtitles.

2 hours later he's got a fully working piece of local software that does exactly what he wants, yet yours is not able to even sort dates correctly. Feel free to download it if you want to see for yourself, I didn't even do any UI tweaks as this was just a tool for him to use:

Linux - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

Windows - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

Mac - https://downloads.blazingbanana.com/whistle-subtitles/unstab...

How can there be such a massive gap in what can be produced?

> frontier LLM development

This seems so wide reaching if it's catching simple things like explaining a paper. Does this also refuse to help with any already developed training pipelines?

I can kind of understand the generation of synthetic data, but nerfing the assistance of training pipelines just seems like a really shitty thing to do.

Can... Can I have this as well?

I imagine it is a bit of a headache, but it is an available option you can choose.

Has to be a personal debit (not credit) card - https://www.gov.uk/pay-corporation-tax/debit-or-credit-card

Paying UK corporation tax can be done with a personal credit card.

Thank you!

Yea absolutely, but man, where to even start, it is very specific.

Fundementally I didn't use any wrappers like unsloth or axolotl, although I have used the latter before a year or two back and it was good, but I needed something very very custom. I also wanted the whole fine tuning pipeline to exported OpenVino model to be seamless.

I heavily leaned on codex, claude and some manual sleuthing around the internet to understand what I needed. I'd played about with QLoRA finetuning with axolotl before and felt most comfortable with that. So I needed to keep everything as stripped down as possible and figured I can just utilise the 3 main huggingface libraries (transformers, peft and datasets) and also bitsandbytes (as suggested by claude to quantize the model to keep this working on my GPU) along with some custom scripts generated by claude/codex (each cross referencing each other) that will do the different stages of the training run.

The next part was the data. Obviously didn't have access to thousands of meetings and associated output documents but I did have a 3090ti sitting there and a codex subscription. So I set about working out what format I needed the data in (many thanks again, to claude/codex) and started generating hundreds of different transcripts, different amounts of speakers, content, tones, subjects, spelling mistakes - like all the different things you could think a meeting would have. Then it's a case of actually generating a good meeting document off the back of the transcripts and creating the "gold standard" that we'd use.

I'm going to gloss over a lot here as I'd rather not detail it as it relates to some propriatary stuff that I had to work through, but you basically pair the transcripts together and run the training.

At the verification stage, there was pretty much 3 things:

1. "just" do some regex string matching to see if there's any of the source transcript key facts in the output to ensure fact preservation. Same with owner fabrication (who said what), I don't want something attributed to someone when it wasn't them that said it and then finally markdown validation.

2. Using codex/claude to validate the transcript and output from the model - I used the latest frontier models, probably overkill for my task, but they were good at the job

3. Finally me going through some actual recordings of myself, groups, meetings and manually verifiying the output

So a fair bit of work, and for context I'm on version 10 now, so it's been a journey!

I think small models have a very good niche for specific tasks. I utilise a fine tuned Phi-4 model (smaller than this one) that fits in about 3.5gb of RAM (not vram) for the document processing side of things for the desktop app I develop (a bit of a shameless plug - whistle-enterprise.com).

If you have a very specific idea for local model use you can find a way to make it work very well, you don't even need to have a graphics card or NPU chip. You just have to be extremely constrained in how it's used. I think as a generic chatbot they're not great, I'd use a hosted SOTA model and I'm a big fan of local LLMs myself.

> I think reality could come down hard here.

I have worked retail before, and to add onto the things you put it was the lack of problem solving for me that was absolutely mind numbing. Sure there were the little "problems" to solve of shelving, stock order, tidiness etc but it doesn't push the brain (and maybe they're done with that part, which is fair), but until you've experienced it I would be very surprised if this person finds retail better than tech.

Yep and when you see it in the wild it stands out like a sore thumb, absolutely no thought into a bit of a unique design or branding.

I have a few live websites built using LLMs and they will just go for default generic templates and colours if there's no vision.

For some reason Logitech have blocked the WebConnect app (for pairing mice/keyboards) from running on Linux.

Scripts in this repo will mirror the site, do the required replacements and OS list update and it'll work on Linux (Only tested on Ubuntu).

Also means that if they retire the page you can still pair your older devices with the unifying receiver. You'll need to use .deb version of Chrome.

> over ten times more than they found in Firefox 148 with Claude Opus 4.6

And how much with Opus 4.7? 5x?

Is it just an exchange for traffic? I run a website that I'm perfectly happy for a single user to not land on themselves with a browser on their device, if they are provided the information that I'm providing or purchase a service through the AI product it doesn't make a difference to me.

Some websites can run only on ads. Is it such a bad thing that they would die off?

I say this as someone that likes the old web and has fun hitting the "surprise me" button on https://wiby.me/ (not affiliated) and browsing the random sites. Just giving an alternative view.

And what's to stop them increasing the fees at any time for "protection"?

I love this and could be very useful in aligning diarization labels. Thanks for sharing

I think tools like this are needed but the current default of "random SaaS bot joins every call and uploads the whole thing somewhere" feels like the worst possible version of it.

The whole point should be that the people in the meeting can actually focus on the conversation rather than half listening while trying to write down enough notes to remember it later. If you're paying people good money to be in a meeting, having them spend half of it doing low quality note taking is a bit mad.

This is basically the reason I've been building Whistle Enterprise (https://whistle-enterprise.com). I'd much rather have something where I choose to record a meeting, process it locally, generate the document and then decide what to keep or delete. I mean yea, it still creates a record so it doesn't solve the legal / discovery side, but at least you're not also adding a random third party into the middle of every conversation.

What I don't understand is how half the comments are calling out how bad the content is, yet it's somehow 4th on the frontpage?

It looks like generic AI slop, the site doesn't even render the headings for their SEO spam "Curated AI Tool Collections by Use Case" section properly and they're half cut off. The images all have the very distinct generic AI hue to them without any attempt of bringing it into a specific style or brand.

Who is upvoting this stuff? Do people not care? Is it just bots gaming the system? Am I an old man shouting at a cloud?

Note that this looks to have been a few days ago:

> 23 March 2026

> I'm about to embark on some months of cold-calling (it will be brutal)

This is something I need to do, have no knowledge about and is definitely going to be harder than building the thing.

> it would be cool to record the calls and feed them to AI for some simple/crude auto-summary which automatically pulls out rejection reason, concerns, interesting points etc

This is very close to what the software I've been building over the last few months is, offline note transcription with summary file generation in a nicely formatted PDF/Docx using local models. Codex is available if you don't want to do the inference on your laptop (I will sort claude out soon as well).

If it sounds like something of interest then feel free to send me a message. More than happy to send you a 30 day trial version.

This is something that interests me a lot. In my own personal experience, ChatGPT is awesome at troubleshooting, it's given me terminal commands that are perfect and use the exact flags needed to identify and then fix the problem.

Why is there this massive disparity in experience? Is it the automatic routing that ChatGPT auto is doing? Does it just so happen that I've hit all the "common" issues (one was flashing an ESP32 to play around with WiFi motion detection - https://github.com/francescopace/espectre) but even then, I just don't get this "ChatGPT is shit" output that even the author is seeing.