Bronze
Back the work
$10 / month
- Your name on our supporters list
- A vote on what we build next
- A thank you in the build log
Open software
fieldkit
A free Python toolbox of the patterns we proved on one small AI desktop. Search your own files, test models, fine-tune them, and ship them. Take only the parts you need.
- Open source
- Python
- Proven on Spark
- Language
- Python
- Install
- pip
- Proven on
- DGX Spark
- License
- Free (Apache 2.0)
fieldkit
fieldkit is a free box of Python tools for building with AI. We made it while doing real AI research on one small desktop, the NVIDIA DGX Spark, and packed in every pattern that worked. You do not have to take the whole box. Pull in just the one part you need and leave the rest.
What is in the box
Eighteen pieces now, and they cover the whole job, end to end:
- Search your own files. Point it at your notes or documents and ask questions. It finds the right parts first, then answers from them. People call this RAG, which just means “look it up, then answer.”
- Talk to a local model. A simple, reliable way to send questions to an AI model running on your own machine.
- Test and grade. Run checks on a model and score how well it does, even using one AI to grade another.
- Fine-tune. Teach a model a new skill by training it on your own examples.
- Shrink to run faster. Make a model smaller so it runs quicker, and measure what you traded away. The trade is called quantizing.
- Publish. Build a clean model card and ship the model to HuggingFace for others.
- Run an agent. Install, set up, and protect an AI agent on your own machine, and route each kind of question to the right model for the job.
- Drive it all from one screen. The parts behind Orionfold Arena: the cockpit itself, a job queue, a spending brake, and a cost ledger. The cockpit can now start and stop the serving model too, with a memory check first and a confirm step before the old model is stopped.
- Remember and recall. The memory layer behind Orionfold Cortex: index your notes, stamp where every fact came from, and check that a rebuilt index can still find its answers.
- Let the model train itself. A training loop where your own tests are the score. The model practices, the tests grade each try, and a hard gate makes sure a worse version is never the one that gets kept.
There are smaller helpers too, for drawing charts, running notebooks, and keeping a tidy record of every experiment you try.
Proven, not promised
Every tool here ran on a real desktop doing real work, not in a slide deck. The same kit built and shipped our open models, and it now powers the Arena cockpit and the Cortex memory layer too. So when you reach for a piece, you are reaching for something that already earned its keep.
Install
pip install fieldkit
Use it
# Will this model fit on the desktop, with this much context?
# fieldkit does the memory math and tells you straight.
fieldkit feasibility llama-3.1-70b --ctx 4096 --batch 32 --dtype fp8
from fieldkit.nim import NIMClient
from fieldkit.rag import Document, Pipeline
# Talk to a model running on your own machine, search your notes,
# then answer using only what was found.
with NIMClient(base_url="http://localhost:8000/v1",
model="meta/llama-3.1-8b-instruct") as gen, \
Pipeline(embed_url="http://localhost:8001/v1",
pgvector_dsn="postgresql://spark:spark@localhost:5432/vectors",
generator=gen) as pipe:
pipe.ensure_schema()
pipe.ingest([Document(id=1, text="The DGX Spark has 128 GB of memory")])
print(pipe.ask("How much memory does the Spark have?")["answer"])
# fieldkit also ships the Orionfold Arena cockpit.
# One command brings it up in your browser.
pip install "fieldkit[arena]"
fieldkit arena up
Specs
- Language
- Python 3.11 or newer
- Install
- pip install fieldkit (from PyPI)
- Proven on
- An NVIDIA DGX Spark desktop, 128 GB of memory
- Covers
- Search, testing, fine-tuning, shrinking, publishing, agents, memory, and a cockpit
- Size
- 18 modules at version 0.31; take one import at a time
- License
- Apache 2.0 (free to use)
Sponsor fieldkit
Back this work with a monthly tier. Your support moves your requests up the list, and Gold or Platinum earns a badge on the roadmap item you back.
Silver
Get a say
$25 / month
- Everything in Bronze, plus:
- Your requests join the priority line
- Early access to new versions
Gold
Move it up the list
$50 / month
- Everything in Silver, plus:
- High priority on your feature and fix requests
- A Gold badge on the roadmap item you back
- Access to Orionfold Discord server
Platinum
Shape the roadmap
$100 / month
- Everything in Gold, plus:
- Top of the line, we work your requests first
- A Platinum badge on the roadmap item you back
- Access to Orionfold Discord server
- A direct line to us and a hand in the roadmap
Need something specific? Send an enquiry from the roadmap.
Keep exploring
Model
Patent Strategist
Offline patent reasoning in ready-to-run files, built with the NeMo toolkit. Nothing leaves your desktop.
Book
AI Research on NVIDIA DGX Spark
Real notes from doing AI research on one desktop. The NVIDIA DGX Spark is a small machine with huge power (petascale means it runs about a quadrillion math steps a second), so you can push local AI further with no cloud needed. Every lesson is backed by code that runs.
Story
My first model on a desktop
I ran my first model on a small computer on my desk. 52 milliseconds to the first word, no cloud, no per-use bill. It felt like a local function, not a service.