Working on Marmot https://github.com/maxpert/marmot recently added support for vector index. My local benchmarks show pretty decent QPS with less than GB of RSS on DBpedia dataset.

Interesting part is that I started off implementing a research paper for indexing and performance was not good enough. I ended up tuning things up for my own use-case and ended up with good enough replicatable RAG store.

I recently started a project called best web extension https://bestwebextension.com/

Basically attempting to modernize a lot of browser extension which I have been using since like a decade ago. Some of them are outdated and unmaintained and some were good for the time.

The project is MIT licensed.

I’m working on Recoil, memory safe, compiled static language (yeah, I know, everybody’s doing it) with Rebol syntax and built-in candies like parsing, finite state machine and rich syntax (that’s given, because of Rebol).

It’s nice to see how well-thought language design can pay off years later, with lower token usage. From entropy POV, Rebol syntax is certainly close to optimal state.

https://codeberg.org/rebolek/recoil

I'm working on GPS tools to help support my current contract. I've found there are no good tools for tracing a route on a map and having a mobile device think it's traveling that route. I'm not just talking GPS coordinates, but speed, direction, motion detection, precise timing between waypoints, being able to play these trips forward and backward, step by step, etc. I'm talking time-travel debugging for GPS applications.

It's early days. I'm not even sure it's possible.

Took a sabbatical off tech marketing recently to focus on my creative work:

An interactive sound sculpture running on an Arduino uno+Pd

Using Mandelbulber as a visual effects layer for my experimental music AV show

I am working on Desiderata (https://github.com/github-of-NMI/Desiderata).

An LLM benchmark for open-weight models only, with secret questions.

The questions are asked multiple times to calculate a consistency score.

The results are available in JSON, containing the hash of the question with the number of correct and incorrect answers, the number of unique answers, and the number of times no answer is given. (Uses \boxed{})

Pagecord. Blog without the slog.

Blogging is more unpopular than ever but Pagecord is somehow growing in popularity. Keen to know what people here think.

https://pagecord.com

(Source: https://github.com/lylo/pagecord)

Building a community-oriented platform for maintaining morphological descriptions of organisms + guides to identify them based off a shared vocabulary.

Been working on it on & off for a couple years, usually taking breaks between refactoring stupid decisions.

https://klados.bio/ Prod site is pretty behind dev branch, basically abandoned normal CI / repo hygiene for the moment

muster and muster-pattern-library. (https://github.com/azide0x37/muster)

an agentic coding scaffold/framework you can reference when building out your next random raspi project. prefer to build around systemd units first; make an idempotent installer script, then put as little as possible custom coding around that.

`impl muster` comes down to: /build out this tool wiring together `patterns` like: C3.dropfolder-trigger; R2.device-binding; C4.lazy-resource-gate

or composite patterns like:

T2R4.device-triggered-conveyor "Bind a physical device event to a bounded ingest job that waits for hot-storage capacity, proves cold-storage capability, stages local work, and hands output to a hot/cold conveyor."

I need to back up a couple hundred DVDs, so with muster I get out:

dvd-ingester T2R4.device-triggered-conveyor

Architecture DVD media becomes ready -> udev rule adds SYSTEMD_WANTS=dvd-rip@%k.service -> systemd runs /opt/dvd-ingester/current/bin/dvd-rip-one /dev/%I --apply -> dvd-rip-one proves DEST_DIR and waits for HOT_DIR capacity -> completed rip moves to HOT_DIR/<run-id> -> dvd-publish-one.timer drains HOT_DIR to DEST_DIR -> publish writes DEST_DIR/.incoming-<run-id> and atomically renames final output

Pipelined; ejects after rip completed. Monitors local disk capacity, retries after NAS comes back online; resumes after random reboot; etc.

Working on https://kapturekafka.dev, a desktop app for Kafka protocol inspection. Think Wireshark or Fiddler, but native for Kafka.

Useful to debug local Kafka apps against any cluster, intercepts the traffic, decodes the protocol. You see interesting (and weird) things when you look at the protocol. Still early, though already useful for local debugging when you know what you want.