00 · A LIFESTYLE CHANGE

Your intelligence.
Forever owned.
Always improving.

ARAIL gathers the understanding. DaC is the press that compiles it into sourced truth. Nucleus mints the printed book — a model you own. AeroLLM runs it on your hardware. PaperAgents puts it to work.

Learn to use AI responsibly and securely, then own a specialized model that runs on your hardware. Humans learn. AI does.
Cloud-optional. No subscription. Local, private, cryptographically yours.

SCROLL

01 · The Vision

Built to educate.

Learn to use AI in the modern world — responsibly and securely — and master any domain with it. QuKaiZen is the harness we're building around AI itself — five parts, one stack. All of it runs on your own hardware — no cloud, no subscription, no data leaving the room.

Served by retrieval — RAG — while a topic is fresh. Baked into the model's weights once it settles. Same subject, kept current.

Built today: the provenance-gated knowledge base, its open API, off-disk inference · Roadmap: the Nucleus bake hand-off, the Fact Steward (automated fact verification)

How it started: on the field's shared R&D — the papers, the techniques, and the engineers who made them understandable. People like Karpathy, whose vision of compiled knowledge set the stage. The thesis traces that lineage.

03 · The Origin

Where it began.

ARAIL is the research lab: autoresearch agents curate arXiv daily, and Buddy — your lab partner — tracks your goals and flags what's worth your attention. Local, zero telemetry, Apple Silicon or Linux from 8GB unified. It's the seed everything else grew from.

PREVIEW

Buddy

ARAIL LAB AGENT

Experiment #47 completed — convergence at 95.2%

You set a goal: "Match the 70B teacher on Linux memory management."

The swarm found 3 new edge cases in NUMA topology since last run. Does that align with where you wanted to take this?

./arail setup one command, two tiers — README on GitHub

Buddy Tunnel reaches him on iMessage, Signal, Slack & more — docs on GitHub ↗

04 · The Company

Declare the team. The agents reconcile.

Small specialist agents for sales, support, ops, and books — declared in TOML, watched continuously, and pulled back to the state you declared when reality drifts. Same stack as the lab; one agent per function the business already does.

team.tomlILLUSTRATIVE
[team.sales]    watch = "continuous"
[team.support]  reconcile = "desired-state"
[team.ops]
[team.books]    # one agent per function
05 · The EngineAPACHE 2.0

400B parameters. 8 GB of VRAM.

Once the book is printed, AeroLLM reads it — streaming the model off your SSD one layer at a time — load, compute, discard, prefetch — so the full weight set is never resident. Speculative decoding stacks on top: a small draft model proposes, the full model verifies in a single pass, provably lossless.

With full credit to AirLLM for the layer-streaming idea — rebuilt in Rust for stability and Apple Silicon (MLX) support.

Faster Throughput

up to 7× — speculative decoding on 70B+ teachers

400B+

Max Model Scale

on 8GB VRAM, layer-by-layer off SSD

~83%

Less Power Overhead

unified memory vs discrete-GPU copies

85%

Cost Savings / Watt

projected — $0.30/W vs $2/W industry standard

06 · The Pipeline

Mine deep. Craft precise. Forge permanent.

Parallel agents mine a frontier teacher across 7 knowledge layers; an adversarial swarm then runs until it exhausts every failure mode — graduation happens when the swarm gives up, not when epochs complete. The printed book is the result: a compact 1–7B Super Skill that knows its domain, owned by you, running local.

Today the pipeline runs deliberately, not on its own — automatic bake from a settled World is the aim.

3B from 500B

Distilled to a model you own

graduates when the swarm gives up, not when epochs complete

3-Gate

Post-graduation certification

general regression · domain mastery · hallucination audit

Ed25519

Nucleus Seal provenance

corpus + teacher + config + training, signed once

<2%

Hallucination rate, hard target

zero fabricated entities · refusal calibration >90%

Nucleus Seal — verify it, don't trust it.

corpus | teacher | config | training → sha256 → Ed25519 ✓ · checked offline — change one byte and it stops verifying

Revocation today is operational, not cryptographic — full scope on /nucleus

07 · The Front Door

Understand all of it.

Learn is the front door for everyone finding their footing — the same press, turned on its own craft. Every concept behind the stack explained plainly, starting with the AI / ML Dictionary and a /what API your tools can call.

Start Here

01 Open the AI / ML Dictionary 02 Pick a World 03 Run the Learn explainer

08 · The Principle

This is not RAG.

This is not prompt engineering.

The Super Skill knows.

Intelligence crystallized into a 1–7B model — sealed, local, near-zero marginal cost per query. It doesn't look the knowledge up. It knows it.

Core Metric

Wisdom per Watt

certified, owned capability ÷ lifetime energy to mint & run it

the point where owning beats renting — and it only moves in your favor

Dictionary →