Joseph Kaiser

Self-hosted systems, model experiments, and web apps.

Background in finance, math, physics, and computer science. Increasingly focused on software engineering. AI tools are linking together and unlocking what UNIX systems have always been capable of. Believe now is a golden opportunity for the everyman to build something new.

The Future I Want to Build.

My goal is to push AI systems as far as they can go, while retaining human at the end of every decision loop. Borrow from military command philosophy: automate everything except the judgment call. By empowering more individuals with powerful systems, we can all benefit from an expansion of intelligence. The PC revolution enabled individuals to calculate models on the way to the meeting with excel. the ai revolution enables individuals to orchestrate vast workflow automations as long as they have compute, rented or owned to burn.

01

Applied model research

Fine-tuning and evaluating small language models on a single RTX 4090. Current project: full fine-tuning SmolLM2-1.7B on elementary math, studying how learning rate, loss masking, and data scale affect performance on 395 problems. I aim to find a benchmark I can get top ranking leaderboard performance and contribute meaningfully to research field in area of consumer grade hardware capabilities.

02

Self-hosted infrastructure

A Dockerized vLLM chat interface that runs on home hardware and serves models over the LAN. An RSS reader on a Raspberry Pi with feed scoring, theme labeling, and newsletter ingestion. I have all these computers I have accumulated in my lifetime, and I have always admired folks who can hook a bunch of random machines together to build a monstrosity. I'd like to create a wired octopus with the personality of a german shepherd and intelligence of von neuman and know-how of elon musk or bill foley other great operators.

03

Human-facing interfaces

Web frontends that surface the right information at the right moment. Lean enough that a person can read, decide, and act without fighting the tool. Always able to go one layer deeper in an intuitive way. I am spending more time than I ever expected on frontend. It is very important to the software I want to use and that enables humans to do important work. It also in a very real sense, is not the ballgame. the UI can obscure real inner workings of a system that live on the backend or in between.

What I build with.

A tighter look at the stack around the work: the languages I reach for, the runtime pieces that keep things light, and the hardware I like deploying on.

Stack Overview

Hover or tap a badge to expand the full name. The stack stays intentionally small: durable languages, compact runtimes, and hardware I can actually own and operate.

01

Languages

Automation, interfaces, and data work without a lot of excess machinery.

02

Frameworks & runtimes

Services that stay legible in production, from containers to local model serving.

03

Deployment surfaces

Machines and network edges where software becomes something I can really live with.

Selected work.

A short list of projects that represent the kind of software I like to make and maintain.

01

ai-chat

A web UI for a locally hosted Dockerized vLLM model, built for private LLM access over a LAN.

  • Docker
  • vLLM
  • LAN
View project

02

myrssfeed

A self-hosted RSS reader for a Raspberry Pi, designed to stay fast, simple, and easy to maintain.

  • Raspberry Pi
  • RSS
  • Self-hosted
View project

03

Current focus

Fine-tuning and evaluation work aimed at pushing a model past baseline benchmark results.

  • Benchmarks
  • Fine-tuning
  • In progress

In progress

Let's connect.

If you are building something useful in web tooling, self-hosted infrastructure, or model systems, I am always happy to compare notes.