01
ai-chat
Self-hosted coding companion on vLLM with streaming, workspace tools, agent harness, and voice.
View projectJoseph Kaiser
Self-hosted systems, model experiments, and web apps.
I spend most of my time coding, building self-hosted systems and experimenting with small machine learning models. Lately that means maximizing the capabilities of my limited hardware stack with clever methods.
Background in finance, engineering and philosophy. Increasingly focused on software engineering and AI/ML research.
Projects
Selected work.
02
myrssfeed
Self-hosted RSS reader on a Raspberry Pi — feed scoring, theme labeling, newsletter ingestion, and a web UI that stays out of the way.
View project03
Current focus
Full fine-tuning SmolLM2-1.7B on AddSub math problems. Learning rate sweeps, loss masking ablations, and data scaling experiments on a single 4090.
In progress
Philosophy
What I keep building toward.
I want to build things that people can get real utility out of, enjoy using, and barely have to think about once they are working. Good software should absorb painful details and stay out of the user's way.
01
Open tools
I want free access to the tools I use, which is a big part of why I prefer FOSS. I would rather work in Python than depend on something closed, and I appreciate the communities around open tools enough that I want to contribute back over time.
02
AI systems
The work centers on pushing AI systems as far as they can go — then designing interfaces that keep a human at the end of every decision loop. Borrow from military command philosophy: automate everything except the judgment call.
03
Efficiency and safety
I care a lot about operational efficiency and safety. I also think customizing your setup matters: a good tool should fit the person using it, like riding a bike that fits or trying to use Excel without Windows and a numpad. The kind of software I want to make should create value by reducing friction, abstracting painful details, and helping people do important work with more confidence and less overhead.
Tools
What I build with.
A compact stack I know well enough to use with intention. These are the technologies I reach for most often, grouped by the jobs they do in my projects.
How I choose
I prefer technologies that stay legible under load, do not punish small-scale deployment, and let me move from experiment to usable software without a lot of excess ceremony.
01
Core languages
These are the defaults I trust for most coding work, from model experiments to browser-facing tools.
- Python
- TypeScript
- Bash
- HTML/CSS
- SQL
02
Model stack
Small-model research is where I care most about efficiency, controllability, and extracting more from limited compute.
- PyTorch
- Hugging Face
- vLLM
- CUDA
03
Applications
I like app layers that stay understandable: thin APIs, simple storage, and interfaces that do their job without drama.
- FastAPI
- SQLite
- Docker
- Static sites
04
Systems
The surrounding environment matters too. I like systems that are local, inspectable, and easy to keep running for a long time.
- Linux
- Git
- Raspberry Pi
- LAN services
Contact
Let's connect.
If you like coding, self-hosted infrastructure, small-model research, or thoughtful software tools, I am always happy to compare notes.
Email is best: [email protected]