Meetups/Infra/2026-01-26

• Cole: developing a database to store photo metadata and labels
• Peter: wrote a Rust app for importing external files into Terraform - http://psychlone.xyz/
  • autoschematic.sh
    • github: https://github.com/autoschematic-sh

Future Topics

• Fourth Monday: paper week (math, science, technology)

Introductions

• [name] - [background]. [goals for meetup, or interests to explore]
• Doug - likes o learn about computers, he's learned so much about infra, he was able to change the CR2032 bats in his server today
• Alex - mostly works on ML things, gonna start a project on env center, hydroponic
• Cole - Linux/infra guy, chemistry project going on (nickel plating a necklace); tried to strip nickel out of bass strings, not high enough purity
• Derek - spec driven development for programming
• Peter - coming for 2 weeks, working on infrastructure-as-code in Rust, imports things to terraform
• Daniel - works on drones, lieks 'em
• Caren - reinforcement learning w/human feedback; lisp compiler for C, high performance computing
• Riley - working on ZFS cool stuff, copy-on-write; when you copy a VM with 20 GB storage only using 14MB; working on a shim for 32-bit apps on 64 bit only distros
• Neha - data infrastructure by day, always wanted to learn random topics so attending infra; finishing up some woodworking like a coffee table
• Riley² - loves a good random project; working on automating their Plex server
• Elan - very interested in VLLM for a project; working on trying to make meshtastic nodes available on the network; a challenge in caddy. Danny has been. looking into meshcore on meshtastic. Go https://radio.noise in the space to see a meshtastic client in action.
• Josh - here learning shader graphics dev
• Zacchae - on Saturday he hosts a Decentralized Web meetup. Set up a live Matrix server for Noisebridge. QR codes posted around the space. Or, go to https://noisematrix.wtf/. You need to create a Matrix account.
• Travis aka "TMC" - longtime infra person, holdover from GGJ. Been hacking on melding Golang with ML stuff. MLX-go is the project. Takes the place of numpy and pytorch
• Trey - here for GameJam stuff
• Xander - from New Hampshire, here to check out cool stuff in SF, somehow landed on NB!
• Lulu - Here to learn everything
• Kevin - addicted to computers and software engineering, has way too many projects on the stack.
• Robert - interested in consumer operating systems, lots of interests! Wants to make a radio one day. Self-hosting AI one day
• Erik - likes to learn things; sometimes he knows things, sometimes it's way over his head, and he really likes that combo.
• Loren - background in platform engineering, scraping, defending NB against scrapers (unironically).

Lesson or Demo

• Read aloud: clarify for meetup. We are taking notes in a riseup pad (or I am--help appreciated, and links). We have meeting notes posted to the wiki. noisebridge.net, search Infra, or Meetups/Infra. (the Infrastructure page has a disambiguation link.)

topic interests expressed

creating a db — 1.5k pictures -- metadata and inventory info.

there exist ML dataset labelling tools... buuut, you should just vibe code it

• Peter - might light to demo their

comparable to Terraform / OpenTofu
git
path based

remind of https://augeas.net/tour.html / https://github.com/hercules-team/augeas
Rust-lib "ron". LSP integrations free. https://lib.rs/crates/ron

vLLM on k8s

Loren on vLLM: implements PagedAttention; without it you wind up with inefficient memory use. PagedAttention allows holding more parallel requests in memory, by chunking requests into memory pages w/a lookup table. PagedAttention has a block size limit, with max token waste 32 tokens. Average wasted memory goes down from 40% on other platforms, to 12%. All the newer libraries are using this technique. Inference depends on key-value cache, and vLLM can improve token throughput by a factor of 2x-5x. vLLM permits use on Pre-Blackwell Nvidia hw. vLLM is production-ready compared to Ollama. Does structure decoding. Supports compute envs other then Nvidia, e.g. Ryzen. See: https://medium.com/@welzin/inferencing-llms-at-scale-with-kubernetes-and-vllm-1504f735947d

• vLLM --

what is it
innovated with implementing PagedAttention for KV cache, essential limited quantity required for inference.

Most labs when the release models, have day-zero support for vLLM

vLLM explanation: https://www.youtube.com/watch?v=5ZlavKF_98U

Demos (cont)

• TMC might do a demo later -- little tool I open-sourced and use every day

tmc/it2 -- https://github.com/tmc/it2

• Derek - SDD (spec driven development) assumed idea is you create a specification for your desired app, e.g. spec-kit MD files. Derek was not satisfied with markdown, instead targeting it with Python. to build the spec using Python. Example of a TicTacToe game, spec defined using Python classes and methods which return the desired behavior for that method. It comes with base classes such as Feature, EdgeCase, ... the intent is to create sharable libraries of spec templates, specs, to promote composability. Gave a demo of TicTacToe implemented in Python as a GUI game, defined using his spec classes.
  • https://github.com/drhodes/libspec

• Karan - Lisp compiler for C - https://github.com/karans4/sysp
  • https://notruefireman.org/articles/sysp.html

• Elan - https://radio.noise/ - Danny and Elan stood up a local meshtastic instance. Used Caddy scripts to inject some JS to auto-connect. Ran into limitations of the software, found some new sw, and have something running in the space. Access the radio directly over USB serial w/ a TCP proxy.
  • http://10.21.0.42:8081/
  • http://10.21.0.42:8081
  • should be stable by next week

local LoRa traffic

Questions

Outro

• Loren -- love to see .noise / noisegarden up, and I want to play with meshtasitc stuff at home.

• Riley - surprised the llm people didn't know about paging, but
• Cadey - cool to learn about vllm, see how it helped with efficiency
• Derek - heartened to see how peter had a typesafe infra for terraform,
• Peter - other demos v cool, spec meta-model v cool, vllm
• Doug - spec driving development, good rigour
• Kevin - liked seeing spec-driving development, saw it in python. Like the level of rigor, looks v promsing, tempting for use, improving maintainability. Mesh doohickey v cool
• Lulu - liked Derek's project, spec dev
• Xander - nice learning about vllm, checking the space out
• Tre - cool, got help, integrating claude & godot
• Travis / TMC - cool to see demos
• Zacchae - last few months: lisp & compilers have subsumed my life, cool to see someone combine them
• Chris - late join, no comment
• Josh - introducted to matrix server, Terraform / OpenTofu
• Elan - saw auto-schema stuff last week, cool. Cool to see ron/auto-ron, really liked strongly typed data structures.
• Eric - people here are using cli for claude code, looks cool, see it operating on multiple servers
• Robert - meshtatic is quite interesting. Might make you own
• Alex - ron was cool. terraform

Questions, Discussion, or Coworking

• [Issue]

For next time

Readings & Exercises

• Readings
• Exercises

Join online

• Try it yourself!
  • Join libera.chat #nb-meetup-infra

https://www.noisebridge.net/wiki/Meetups/Infra