deekay

world's greatest deployer

[deekay @flexdinesh]

Hey now that mutuals can see my tweets, here's what I've been up to in the last few years and especially this year.

1. I got deep into distributed systems. I built an on-prem deployment infra from the ground up in my home. It's a 3-node HA virtualised cluster. I run a few things on it - a 9-node Kubernetes cluster, a multi-node object storage, an ingress gateway, a backup file system storage, etc. I host many apps for my family and friends on this cluster. This also helped me earn my Google Cloud Cloud Architect certification.

2. I built a few CLI tools that I use on an everyday basis:
- gitsy (git pull all repos in a dir)
- tokeninsights (analyse token consumption metrics)
- ssh-drop (drag and drop screenshots for a remote ssh agent session)
- checksy (check network connectivity)
- cbox (WIP - a way to sandbox cli harnesses using docker)

3. I built a few server tools:
- key-keeper (a simple auth gateway for Caddy)
- backup-sidecar (not oss, i use this for backing up data from kubernetes services to a NAS storage on schedule)
- models-dev-rss (rss feed for models(dot)dev AI provider data)

4. I built a few websites:
- models-dev-website (a fork of the original models(dot)dev website to make page loads and search faster)
- tools-wiki (a collection of wikis for various popular cli tools)

5. I started writing again and published a few posts in the last year

6. I gave a few talks at SydAI this year

I can't recollect everything. It's been an exciting year of building stuff. I am very excited about what I am building now. Hopefully I will be able to talk about it soon.

(links in thread)

[deekay @flexdinesh]

Arto is my favourite way to learn how something works. I gave my agents an artifact platform and they write blogs about other blogs or oss repos.

I am reading about how portless works in the morning today. @ctatedev cooked with portless.

https://t.co/DoZsWXWY43

[Malte Ubl @cramforce]

Vercel presents a @shadcn and @shuding production.

Seeing the clarity of vision on a space that was considered largely solved has been nothing but inspiring.

The first UI library designed for agentic composition of its primitives and with agentic use as first-class feature. For example, every UI SDK app exposes a WebMCP for ultra-efficient agentic usage of the application.

[deekay @flexdinesh]

the bots found me

[deekay @flexdinesh]

I give this tip to everyone I talk using Claude Code. Set the auto compact limit to 400k. Both your usage and output quality will be better. There's a reason Codex defaults to 272k window limit for frontier models.

[deekay @flexdinesh]

@AndaristRake This was my biggest frustration with Claude Code as well. T3 Code solved it for me. It's still Claude Code underneath. It just orchestrates the sessions via Claude Code.

[deekay @flexdinesh]

I give this tip to everyone I talk using Claude Code. Set the auto compact limit to 400k. Both your usage and output quality will be better. There's a reason Codex defaults to 272k window limit for frontier models.

[Shayan @ImSh4yy]

If you're using Claude Code, set the auto-compact window to 400k and your usage will drop significantly.

[Ryan Dahl @rough__sea]

Check out my talk at Cloudflare Connect on celld.

celld let’s you self-host workers, durable objects, queues, and workflows using only an object store bucket as a backend.

We’re in the process of merging the ideas in celld into the open source workerd. More to come on the new workerd soon!

https://t.co/z9ovyhAEiV

[Brooke Jamieson @brooke_jamieson]

My favourite pre:Invent launch so far - AWS just launched an AWS emulator, and it's built & maintained by thousands of AI agents!
Get started at: https://t.co/FsU8SQZd6J

[NVIDIA AI Infrastructure @NVIDIAAIInfra]

Announcing two open-source projects for building and operating managed Kubernetes services on GPU infrastructure: NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR).

Both are components of the NVIDIA DSX AI Factory Platform. NVKE is the control plane for running managed Kubernetes service and NVKR builds the OS images, container images, and packages the managed clusters depend on.

Available in public preview, get started: https://t.co/c1YuOc7Wjo

[Thinkmill @thethinkmill]

We partnered with Lyka to modernise its platform and bring its new cat food brand to market. Our embedded team worked across four engineering squads, building customer experiences and a shared design system for both brands.

Check out the case study 👇
https://t.co/TQWegflFrx

[deekay @flexdinesh]

@samwillis the best tool for the job now is the one model knows well

[deekay @flexdinesh]

@cramforce @andrewqu I take it that you're running HA on k8s? In which node are you plugging your zigbee/etc dongles lol? I tried that and kept moving my workloads between VMs so got an rpi and HA has been running there happily for a few years now.

[Simon Willison @simonw]

Something I enjoy about these prompting guides is that it's VERY clear from reading them that Anthropic are just as surprised by the idiosyncrasies of their new models as we are

[deekay @flexdinesh]

@stephenhaney The spacing of chevrons in dropdowns. All current models universally refuse to build it right without feedback.

[Kyle Galbraith @kylegalbraith]

it can not be overstated how big of a deal this is.

https://t.co/91Hy77hgKd

[deekay @flexdinesh]

@josevalim the problem with third party solutions is it will require installing a github oauth app in org accounts which many orgs won't be comfortable unless it's from a well known application like slack. email filters are the lowest friction / highest value option.

[deekay @flexdinesh]

@LukeParkerDev I’ve been doing the same since gpt 5.5. Usually I switch to the newer model once I hear enough _this is awesome_ and the FOMO will kick in.

[deekay @flexdinesh]

@craigweiss i've built one for me on top of t3 code. what does that make me?

[deekay @flexdinesh]

@localhost_5173 i have 3 mygo apps going on. it's great except now my dev flow can't be fully remote in a headless ssh machine. need to figure something out for that.

[deekay @flexdinesh]

@tomlienard @vercel Congrats!

And did you just throw in Fluid compute into the tweet list like a weekend project that you shipped? It completely redefined Vercel's compute layer. I still remember Lagon too.

[dax @thdxr]

point your agent at this and build whatever https://t.co/eZhSqXxBf6