QuKaiZen · One developer's working library, open to visitors
Knowledge Compiled.
Models Streamed.
Local Agents.
One developer's working library, kept open for visitors. DDaC compiles a subject into sourced context you can curl. QueueLLM reads a model off SSD a layer at a time, so memory holds a few layers rather than the whole model — 43 tok/s on a Qwen2.5-7B. ARAIL and TEAM run the whole thing on one machine.
ARAIL + TEAM
The lab does the reading. TEAM acts on it.
ARAIL is the lab: it collects a subject, stores it, and measures what it holds — with a local Buddy to learn beside. TEAMis the crew on top: it teaches, explains, answers, and maintains — local agents running on the developer's own machine today.
DDaC + Worlds
Sourced context, with a dial on it.
Point DDaC at any subject and you get a World: a body of sourced context your agents reason from — infrastructure-as-code, but for what they know. You choose how much of it to read: terse for a glance, balanced for understanding, expansive for the full story.
QueueLLM + Nucleus
Don't shrink the model. Stream it. Then mint your own.
QueueLLM streams a model off SSD instead of holding it whole in memory — measured at 43 tok/s running Qwen2.5-7B on an M5 Pro.
Nucleus is the press behind it: a pipeline that distills a settled World toward a small model you could run on your own desk. The pipeline is built end to end, and today it runs deliberately, by hand.
→ we aim to· automatic bakeDDaC generates each World. ARAIL stores and measures it. TEAM executes on it for you. All three run on QueueLLM, and Nucleus can bake it into a model you own.
Same term, three depths
Each entry in a World carries three fields: a short line, a full definition, and a worked example. Slide the lever and read one real term at each length — terse when you have a minute, balanced when you want to understand, expansive when you want the whole story. Each depth is a field the World already holds — the same fields the docent, each World's built-in guide, and the open API answer from.
Context Window
The context window is the hard cap on how many tokens (prompt + generated output) a model can process in a single pass. Everything outside it is invisible to the model, which is why long documents are chunked and agents need external memory. Larger windows cost more compute and KV-cache memory, roughly with length.
terse = short · balanced = definition · expansive = definition + example
The World decides what is true. You decide how verbose — terse, balanced, or expansive — and the docent and the API answer at that length.
The World Press · DDaC + Worlds
The context your AI runs on.
A World is not the finished product — it is the ground you build on. It lays down the definitions and the base structure of what's to be: the shared, fully-sourced vocabulary you and your agents develop from. You don't start from a blank page, you start from a settled foundation.
Working on something ecological? Enter a focused World like Ecology for a tight, specific base — or a higher-level one like Biology to develop with more surrounding context. You pick the altitude; the World sets the ground.
Any subject can be built into one, which is the point rather than the shelf it sits on. A World narrows what your agents reason from to a fixed, sourced vocabulary instead of whatever the weights recalled or the open web returned that morning — and narrowing the ground is what pulls the same question toward the same answer each time it is asked.
A goal implies a theme. Agents research it while a provenance gate checks every source: every entry cites its origin, every cross-reference resolves inside the World, and every category is declared before it ships.
QuKaiZen
editorialThe World you are standing in. QuKaiZen's own vocabulary — the organs, the lifecycle a fact travels, the mechanisms that compile it — compiled and gated by the machinery the terms describe.
Enter the World →
AI / ML
editorialThe original. The whole discipline, from attention and transformers to RAG and agent memory.
Enter the World →
Astronomy
immersive · cosmicThe observable universe, defined. Parsec to the cosmic microwave background, with the stars out at full tilt.
Enter the World →
Mathematics
immersive · formulaeThe famous formulas and the results that changed the subject, Euler's identity to Gödel. The blackboard never stops writing.
Enter the World →
Quantum
immersive · quantumThe physics of the very small, told honestly. Superposition, entanglement, qubits, the myths stripped out.
Enter the World →
Biology
immersive · bioThe science of life, from the double helix and the cell to evolution and the immune system, myths corrected.
Enter the World →
Ecology & Ecosystems
editorialTrophic cascades, regime shifts, and metapopulations — ecosystem-level ecology told honestly, pop-myths corrected.
Enter the World →
World Models
editorialBayesian surprise vs. prediction-error curiosity vs. intrinsic motivation — precision vocabulary for wiring a curious, world-modeling agent. WorldLLM is the flagship method.
Enter the World →
Neural Networks
editorialThe substrate everything modern rests on — from the perceptron and backpropagation to transformers, residual connections, and the failure modes every deep learning practitioner hits.
Enter the World →
GeoAI
editorialThe onboarding world for the humanitarian mapping mission — satellite imagery, MapSwipe and OSM, the QuKaiZen stack, and the zero-egress runtime, strung along the loop a prediction travels.
Enter the World →
Nucleus Pipeline
editorialThe bake pipeline, defined term by term — seven stages from ingestion to the cryptographic seal, and the swarm that runs until it gives up.
Enter the World →
Hello Kitty World
editorialA world for fun. The first domain of Charlotte's World — kindness, cleverness, and a little moonlight: talk past Kuromi by day, race the Glooms by night, befriend the ghosts to escape the manor.
Enter the World →
The Squishy World
editorialNewly dreamed. The second domain of Charlotte's World — a cozy, pastel realm of plush, squishable friends that runs on softness, comfort, and a warm hug. Her own creatures, growing as she dreams them.
Enter the World →
Grade 4 Mathematics
editorialThe third domain of Charlotte's World — her actual Grade 4 Mathematics, the five Common Core strands, written so a nine-year-old can read them. The first World with a level on it.
Enter the World →
Grade 4 English
editorialThe fourth domain of Charlotte's World — her actual Grade 4 English, grammar to writing, the sibling of Grade 4 Mathematics at the same altitude.
Enter the World →
Because each term is indexed on its own, an agent looks a fact up on demand instead of carrying the whole dictionary in its context window.
Learning Lab · ARAIL + TEAM
The Press publishes worlds. TEAM is where a world becomes yours.
ARAIL · the bench
Point the lab at a subject and it collects the literature, stores it, and measures what it holds against the goal you set — accuracy, cost, latency, drift. Buddy, a local companion, works the bench beside you. It runs on your own hardware, offline by default.
TEAM · the crew
Teach · Explain · Answer · Maintain
Four verbs, one crew of local agents: teach it a rule once and the correction becomes something you keep. Sharing moves rules, never data — a categorizer travels, the records behind it stay home.
Underneath: PaperAgents — its shipped pactl validatecontract gates TEAM's plans today.
Control with DDaC. Collect and build with ARAIL. Ship with TEAM.
The Bindery · QueueLLM + Nucleus
Knowledge, pressed small enough to run on a desk.
Don't shrink the model. Stream it. Memory is the expensive resource and the one you cannot always buy at a sensible price; disk is neither. So QueueLLM reads a model off SSD one layer at a time — load, compute, discard, prefetch — and the full weight set is never resident. Measured at 43 tok/s and 3.13× mlx_lm running Qwen2.5-7B at bf16 on an M5 Pro.
Nucleus distills a settled World toward a compact model that runs where you are, and measures itself in Wisdom per Watt — capability kept, per watt spent minting and running it. The seal records one thing: where a model came from. The pipeline is built end to end, and today it runs deliberately, by hand.
→ we aim to· automatic bake from a settled WorldApplied · pointing the stack at a good cause
The first place it goes to work outside the lab.
Everything above gets built and exercised inside ARAIL. At some point a stack has to leave the workshop and do something for someone, and the interesting question is what you point it at first. This is that: not a demonstration arranged to flatter the machinery, but a real problem picked because the machinery happens to fit it.
When a disaster happens, someone has to map it before responders can move. The OpenStreetMap community does that work, and the machine learning that speeds it up runs on inference somebody has to pay for — which means cost is what rations how much gets mapped and how fast.
X-GeoAI is this stack aimed at that problem: vision-language models that compare pre-disaster and post-disaster imagery and return plain-language claims with the evidence behind them, zero-egress by design, so imagery never leaves the machine it lands on. DDaC declares the run and pins the sources, QueueLLM streams the model off SSD, and every inference mints a signed receipt saying what ran and on what.
The piece worth building next is the one that changes the economics: a scheduled worker, shaped like SETI@home, that a volunteer runs on the Apple Silicon they already bought. The batch work and the inference both fit on an edge node with a small footprint, and they run there safely — the imagery stays put and the node hands back a receipt, not a dataset. The machine is idle most of the day. Point it at a crisis map overnight and the inference bill for humanitarian mapping stops being the thing that rations it.
The cycles are the smaller half of the contribution. A model that says “building destroyed, roof collapsed on the north side” is only useful if someone can say whether that is right — and why. That verdict is the part no volunteer network has ever been asked for and the part this design is built around: a person confirming or overturning a claim against the evidence it cited, and the reasoning travelling back as something the next model can be taught from.
→ we aim to· the volunteer worker networkIt is the first of these rather than the reason for them. What the stack is actually for is whatever subject someone needs it on — the parts that make this work are a World, a model small enough to run where the data is, and a person whose judgement is worth keeping. Crisis mapping is where that combination earns its keep first.
For builders
An open API, not a login.
$ curl qukaizen.com/dac $ curl "qukaizen.com/what?world=math&term=quadratic-formula"
/dac is the framework manifest — every World, every route, in one response. No key, no account.
A Personal Wing
A ten-year-old's library, too.
Charlotte's World of Imagination is the founder's own family experiment, played from the Break Room arcade — Hello Kitty and a realm of squishy plush friends, compiled through the identical gate as every World above. Her Grade 4 Mathematics and Grade 4 English are the newest arrivals: her actual schoolwork, sourced to the Common Core standards and compiled the same way — the World factory pointed at a fourth grader's homework.