hxii

It's interesting to see use-cases like this "in the wild" (as opposed to some hobbyist group trying stuff out).

With a little investment, I think Meshtastic/Reticulum could do wonders for communities like this. With Argentina being largely flat, even putting up a relatively short mast on the roof of a house should give it a good enough range (~10-30km).

Or wallhop.io

So if someone could physically gain access to your server/computer and press "delete" –– would that mean your service is not self-hosted?

Location: Israel

Remote: Yes

Willing to relocate: Probably not.

Technologies: Python, AI/LLM/MCP, Grafana, N8n, GitHub Workflows, Okta, Terraform, some Ruby and some Crystal

Résumé/CV: https://hxii.github.io or https://0xff.nu

Email: cv <at> glushak <period> net

I'm a polymath with ADHD which means I'm very interested in many diverse subjects and often can connect the dots between otherwise unrelated subjects. I have recently made a fun, toy project called koi.rest, picked up model benchmarking and training for fun (and personal profit), taught my wife to use Linux and LLMs and got her to learn Python.

I like making fun, interesting and novel projects for people, automating stuff and finding solutions to annoying problems.

The curve on both ends is very brief (20ms, so 40ms total) and is more to deal with any abrupt clicking.

The audio was sourced from pixabay, and is not looped.

The start and end points have gain faded in a curve, to avoid any clicking especially since the audio samples are not seamless.

Previously a second buffer was used for crossfading, but it doubles the memory consumption with very little benefit.

I was toying with the idea of synthesizing audio from smaller samples directly, but that's an idea for the future :)

Thank you for the feedback! Will definitely add that as an option.

Vibe coded doesn’t necessitate bad performance. As I’ve replied to others – during testing I didn’t get any lag across three different devices that I own, so I just didn’t have a chance to catch this before.

That’s not to say “it’s a you problem”, but rather that I currently don’t know under which circumstances the lag appears.

Thank you for the feedback! It’s definitely on my project task list.

So far the options I thought of were to fade or scale the koi as they enter or leave the pond. I will try to see what will work visually here.

Vibe coded or not, I was honestly trying in on two desktops (and two browsers) and a mobile phone and didn’t get any lag, which is why I couldn’t have caught this before.

It might be some other combination that might contribute to whether there is or isn’t lag, and I’m going to try and find what it is :)

Thank you for the feedback!

During my work on it I’ve been testing it on both my Linux box as well as my MacBook and haven’t encountered lag. I’m not meaning to say “it works on my machine” by this, but simply that I couldn’t have caught this before.

I will definitely address the lag issue, and find a more “graceful” way for koi to enter and leave the pond.

About interactivity – I was thinking about this both during the development as well as just when I had the idea and the intention was always to not have any, simply because in my own experience we are always doing “something”, and this was a way to get a little bit of the opposite – to do nothing.

But I can see how some micro-interactions could be useful. A sort of a “popping bubblewrap” action.

Hey! That’s the base, correct. Three are always there so if someone joins and nobody else is using it – they’re never alone.

In essence, when you someone joins the counter goes up and another koi will spawn locally.

It’s a bit nerfed from my initial idea to sync the movement of each koi, but I might do that in the future.

It looks like the way I deployed (Railway) it causes `429 Too Many Requests` to be returned, so I’ll definitely have to look into that. But it’s 1:21 AM so tomorrow.

Thank you for letting me know!

Location: Israel

Remote: Yes

Willing to relocate: Depends. Maybe.

Technologies: Python, AI/LLM/MCP, Grafana, N8n, GitHub Workflows, Okta, Terraform, some Ruby and some Crystal

Résumé/CV: https://hxii.github.io or https://0xff.nu

Email: cv [at] glushak.net

Recently picked up model benchmarking and training for fun (and personal profit) – https://0xff.nu/hexbench

I spent a decade in support, slowly automating myself out of every role until the tools I built became more valuable than the tickets I closed (at least to me). Now I write Python, creating internal tooling, LLM integrations and automations.

Current projects: hex (https://0xff.nu/hex), an AI executive assistant that manages my calendar, tasks, and Obsidian vault because my ADHD brain certainly won't, and sixwhyo (https://6yo.dev), a code reviewer powered by a 6-year-old's logic. Now also exploring iOS development to create an app I’ve been dreaming about.

I like simple and robust solutions over overly clever architecture, and am looking for a remote-first engineering role, not a support role with "engineer" in the title.

I’ve been running this level of models through a personal benchmark on a RX6750XT.

I am hoping to get to the point of being able to run a personal assistant agent with tools at its disposal.

The results are largely the same, with some outliers.

Results can be seen here: https://0xff.nu/hexbench/

In my brief testing, it did about as well as Qwen3.8-4B-Distill, and LFM2.5-2.6B overtook both.

It did an absolutely terrible job at generating CSS, where I instructed it to finish implementing a bright and dark theme based on a palette through the use of `color-mix()` and it just went ahead, removed everything I pre-added and replaced it with hardcoded hexadecimal color values.

Interestingly, in my own benchmark and testing (in the hopes of finding a good-enough local model to run a personal assistant agent), Ornith-1.0-9B was worse than Qwen3.5-9B which according to their scores should've been reversed.

I will definitely pass Ornith-1.5-9B through the gauntlet as well!

Location: Israel

Remote: Yes

Willing to relocate: Depends. Maybe.

Technologies: Python, AI/LLM/MCP, Grafana, N8n, GitHub Workflows, Okta, Terraform, some Ruby and some Crystal

Résumé/CV: https://hxii.github.io or https://0xff.nu

Email: cv [at] glushak.net

I spent a decade in support, slowly automating myself out of every role until the tools I built became more valuable than the tickets I closed (at least to me). Now I write Python, creating internal tooling, LLM integrations and automations.

Current projects: hex (https://0xff.nu/hex), an AI executive assistant that manages my calendar, tasks, and Obsidian vault because my ADHD brain certainly won't, and sixwhyo (https://6yo.dev), a code reviewer powered by a 6-year-old's logic. Now also exploring iOS development to create an app I’ve been dreaming about.

I like simple and robust solutions over overly clever architecture, and am looking for a remote-first engineering role, not a support role with "engineer" in the title.

I’m wondering if they did anything to address the DSML tool calls leaking. Has been an issue with both Flash and Pro so far.

I have a question if you don’t mind – do you also have to kill AWDL to prevent lag spikes during a streaming session?

I'm working on adding better Obsidian support to my personal SSG – Hajime [1], in an effort of making it a somewhat solid Obsidian Publish alternative.

The main point is adding relationships between the entries, as that's the bread and butter of Obsidian.

[1]: https://0xff.nu/hajime/

The cost of design, build and marketing do not triple the final price by themselves.

The kickstarter was successfully funded, yielding in over $3.3M. The device back then was $149 at 20% off, with $169 being the full price. That's still not $193.

Again, this is just my personal opinion, but even taking everything into account and factoring in a nice profit margin, being OVER $100 is very, very hard to justify.

I understand the rationale, but it looks like they decided to go the Sisyphean route on this.

ESP32-C5 ~$10

ESP32-P4 ~$5

Display ~$10

Case is maybe 90 cents (let's round to $1) if you print out of ABS.

The rest of the components? Let's say another $15 worst case.

~$41 or around that much? Sure, but definitely not ~$193.

Thanks! It took me some time to get to something I like, with the original idea being made 6 years ago: https://themes.bludit.com/theme/u

Thanks! You made me realize I completely forgot to mention memory in my post .

For running memory, hex maintains a 10 message full-length context window, with the rest being summarized when the threshold is being crossed (15 messages) by a separate LLM call and stored in vector storage (Qdrant). Aside from that, observational memories get constantly stored in Qdrant as well for convenient retrieval and embeddings are calculated locally via Qdrant's FastEmbed.

Preferences, facts and things like that are store in SQLite.

Location: Israel

Remote: Yes

Willing to relocate: No

Technologies: Python, AI/LLM/MCP, Grafana, N8n, GitHub Workflows, Okta, Terraform, some Ruby and some Crystal

Résumé/CV: https://hxii.github.io or https://0xff.nu

Email: cv [at] glushak.net

I spent a decade in support, slowly automating myself out of every role until the tools I built became more valuable than the tickets I closed (at least to me). Now I write Python, creating internal tooling, LLM integrations and automations.

Current projects: hex (https://0xff.nu/hex), an AI executive assistant that manages my calendar, tasks, and Obsidian vault because my ADHD brain certainly won't, and sixwhyo (https://6yo.dev), a code reviewer powered by a 6-year-old's logic.

I like simple and robust solutions over overly clever architecture, and am looking for a remote-first engineering role, not a support role with "engineer" in the title.