ARAIL — EXPERIMENTS & EDUCATION

A rail for AI experimentation
& education.

ARAIL ships as a blueprint. Take an agent and an IDE and tailor / set up yourself. A Python runtime, AeroLLM, a starter model, and supportive agents — including Buddy. Everything you need to start a session offline, on a single box.

If you have a smaller machine, choose Minimalist mode. Else set up Maximum mode to build out all the functionality: a Rust router, AeroLLM, an agent swarm, AutoResearch, frontier teacher models.

QuKaiZen · ARAIL · v0.4 · airgapped

02 — ORIGIN

Why I built this.

First there was Buddy. Then a place for him to live.

I wanted to learn AI — properly. I was curious how the models were built. I didn't know where to begin, so I started with a companion: Buddy. Then I built an environment for him. That environment became ARAIL.

Then I found agentic AI, then autoresearch — and couldn't get enough. The lab needed to be pluggable, and I had to be able to see what was happening inside it.

03 — THE RAIL

AutoResearch AI Lab.

The name is the goal. A rail — straight, fast, repeatable — that you mount things onto. The rail is the scaffolding. What rides on it can change month to month — the operating surface stays the same.

agentsautoresearchdual-modelknowledge
Mission control — quick actions, live services, and your current mission

THE LAB — MINIMALIST CORE · MAXIMOST REACH

MINIMALISTMAXIMOSTBUDDYAGENT · PERSONALITYPYTHON RUNTIME3.10+AIRLLMlayer streamingSTARTER MODELQwen3-8B · 4-bitSKILLS · PKBonboarding · observe · setupRUST ROUTERqkz · routing · CLIAEROLLMSSD-streamed · 70B-400BSWARMinterrogator · adversary · evaluatorAUTORESEARCHrubric evolutionTEACHER700B+ frontier · streamedSKILL PACKScuration · model-building · researchBUDDY TUNNELhybrid only · gateway requiredGATEWAYTelegram · Slack · Discord · SignalBuddy — personality coreMinimalist — always onMaximost — opt-in expansionHybrid only

05 — DEFINITION

Three keystrokes to a local AI lab.

A clone-and-run research bench. Dashboard, chat, autoresearch loop, knowledge base, agents — server-rendered, no SPA, no telemetry. Default mode is airgapped. Zero network calls until you flip LAB_MODE=hybrid.

$ git clone cdarnell/arail
  cloned to ~/arail
$ ./arail setup
  ✓ venv created  ✓ deps pinned  ✓ portal ready
$ ./arail start
  ▶ serving on 127.0.0.1:8080
airgapped  ⬤ buddy · sre · researcher
$ open http://127.0.0.1:8080
Lab docs — the AI dictionary and runbooks, in-product

04 — COST POSTURE

Pay for frontier models. Just not by accident.

Simulate before you spend.

LOCAL MODE — AIRGAPPED

$0

Iterate freely. Run a thousand experiments overnight. Zero network calls until you say so.

HYBRID MODE — METERED

$$

Every cloud call is logged with provider, model, and tokens. Same surface. Same agents. Real receipts.

06 — TIERS

Two tiers, one surface.

TIER · STARTER · MINIMALIST

The lab

  • Dashboard
  • Chat — local + cloud
  • Autoresearch loop
  • Knowledge Base
  • Agents — Buddy · SRE · Researcher
  • Airgapped by default

TIER · OPERATOR · MAXIMUM

The operator's console

  • everything in Minimalist +
  • Admin · Docs · Notebooks
  • 405B local inference via AeroLLM
  • Cloud SDKs · provider routing
  • Still airgapped unless you flip LAB_MODE=hybrid

BUDDY TUNNEL

Hybrid only — gateway required

The dashed branch in the diagram. Telegram, Slack, Discord, Signal cleanly; iMessage and WhatsApp through bridges. Needs internet and a gateway. Airgapped mode blocks the tunnel by design — see docs/BUDDY.md.

07 — PLUGGABILITY

What plugs into the rail.

Four mounting points. Swap any one. Same protocol, same activity stream, same token accounting.

Agent Control Center — forge an agent from a prompt, deploy without a restart

08 — COMPARE

Same prompt. Two minds. One screen.

Local vs cloud. Small vs large. Base vs fine-tuned. Real comparison, not benchmarks.

parse_goal> explain the bias-variance tradeoff like I write firmware

QWEN-3-8B

local · MLX

Bias is your model's stuck-at fault — it always answers a little wrong. Variance is marginal noise — different inputs ring differently. You tune capacity to balance them, same as a low-pass filter…

184 tok·3.1s·$0.000

CLAUDE-OPUS-4.1

cloud · anthropic

Think of bias as systematic miscalibration baked into your model class, and variance as jitter from the specific training samples you happened to draw. They trade off through model capacity…

212 tok·1.8s·$0.024

09 — CACHING

Pay for the context once.

Every call ARAIL makes to Claude opens with the same fixed context — the system prompt and instructions that don't change between runs. ARAIL tags that stable opening with cache_control: ephemeral, so Claude keeps a ready-made copy of it. The next call that starts the same way doesn't pay full price to re-read it: those tokens bill as cache_read— a fraction of the cost of fresh input — and only the new, changing part of the request is charged in full. This is the provider's cache at the API edge, not the model's per-request KV-cache. See prompt caching.

RESEARCHER

clear win

Threads an identical system context across 3-5 calls per run. Calls 2-5 hit cache_read instead of fresh input. Repeatable, predictable, the load-bearing case.

stable prefix · repeated dispatch

CHAT

kicks in at depth

A ~1.2K-token single-shot chat sits below the Sonnet 4 floor — the client omits the marker, no penalty, no win. Around turn ~3 the transcript grows past the floor and caching engages.

marker omitted below floor

⬤ THRESHOLD-AWARE · CLAUDE SONNET 4 FLOOR: 2048 TOK · LOCAL & AIRGAPPED PATHS UNAFFECTED

10 — OBSERVABILITY

If you can't see it, you can't trust it.

Every agent action — timestamped, attributed, token-counted, streamed live.

Activity stream

tail -f · live

14:02:08 buddy parsed goal → p95_latency < 400ms 412 tok
14:02:11 researcher spawned experiment exp_0014 qwen-3-8b · MLX · 88 tok
14:02:24 researcher baseline measured · p95 = 612ms 1.2k tok
14:03:01 sre ⚠ vram pressure 88% · throttling to bs=4
14:03:48 researcher applied kv-cache patch · p95 = 387ms ✓ target hit · 3.4k tok
14:03:50 buddy notes written → knowledge_base/exp_0014.md 204 tok
Autoresearch — every experiment is a git branch; the loop runs your program.md

11 — MEMORY

Knowledge base — the gem.

The lab gets smarter the longer you use it. Drop data in. Agents drop data in. Relevant pieces get RAG'd back into context — for you and for them.

KNOWLEDGE_BASE/

RAG · retrieve · augment · generate

YOU

papers · notes · code

BUDDY

distilled summaries

RESEARCHER

experiment notes

SRE

crash transcripts

The knowledge base — one tree feeding the wiki, notebooks, and agents

12 — CONSTRAINT

The hardware wall.

Even a $5,000 GPU still settles.

The compromise: < 100B parameters · small context · loud fans · hot rack. That wall is real. You can buy your way closer to it. You can't buy your way through it.

So I stopped trying to buy through it. I built around it.

WHAT YOU ACTUALLY WANT TO RUN

405B

WHAT FITS — AT SMALL CONTEXT

~70B

13 — INFERENCE

AeroLLM — built around the wall.

Layered inference for MLX and CUDA. Same idea AirLLM proved — stream weights in, compute, evict — rebuilt for unified memory and quiet enclosures.

# INSPIRATION

▸ AirLLM

  • + proved the idea
  • unstable on long runs
  • no MLX support
  • CUDA-only · 3090 sounded like a leaf blower

# THE ANSWER

▸ AeroLLM — same idea, made for laptops

  • MLX + CUDA
  • stable on hour-long runs
  • ships with ARAIL Maximum

Thanks to the AirLLM team for charting the path.

14 — HARDWARE

We meet you where you are.

No pretending. Minimalist runs almost anywhere. Maximum works best on a decent gaming or MLX system (only MLX tested with AeroLLM).

MINIMALIST

cpuany modern x86_64 / Apple silicon
ram16 GB
gpuoptional
disk20 GB
osmacOS · Linux · WSL2
netnone required

MAXIMUM

cpuM-series MacBook Pro · serious gaming rig
ram64 GB unified / 64 GB sys + 24 GB VRAM
gpuRTX 3090 / 4090 · M3 Max / M4 Pro+
disk500 GB NVMe
osmacOS 14+ · Linux
netrequired only for hybrid mode

15 — NEXT FRONTIER

Project Nucleus — a model that learns from your knowledge base.

Nucleus reads the world's published research on distillation and efficient fine-tuning, then applies it to a base model + your KB. You get a smaller, sharper, you-shaped model.

01 · INGEST

knowledge_base/

papers · notes · experiments

02 · SURVEY

techniques

distillation · LoRA · DPO

03 · TRAIN

recipe

picked, not guessed

04 · SHIP

you-shaped model

smaller · sharper · local

⬤ Prototype runs ARAIL by default — a 3B-parameter AI specialist trained by Opus themselves, built first-hand by Nucleus with the latest AI research from sites like arXiv :)

16 — AVAILABLE

A QuKaiZen product. Built the kaizen way.

Available by request.

MINIMALIST

⬤ AVAILABLE

Dashboard · Chat · Autoresearch · Knowledge Base · Agents. Airgapped by default · hybrid opt-in. Runs on any modern laptop.

MAXIMUM

⬤ AVAILABLE

Everything in Minimalist + Admin · Docs · Notebooks. AeroLLM · 405B local inference. Cloud SDKs · provider routing. Serious rig or M-series MacBook Pro.

NUCLEUS — ADD-ON

⬤ PREVIEW

Fine-tuning pipeline. A smaller, sharper, you-shaped model distilled from your knowledge base. See /nucleus.

ACCESS

By request · info@qukaizen.com

Not open source · Pull requests not accepted · Source and Access available to friends and family. Open to all soon.

Request access →
Dictionary →