ARAIL · Auto Research AI Lab
An autoresearch lab.
It is not an app you log into. It is a blueprint you assemble — a runtime, a streaming inference engine, a router, and a swarm of agents — and at its center, a knowledge base that everything writes to and learns from. Nothing leaves your hardware.
At a high level the lab does three things: it runs research on your goals, it builds models, and it curates what it learns into that central knowledge base — which you can simply chat with.
01 · Autoresearch
Name a goal. The lab runs the scientific method on it.
Autoresearch is the reason the lab exists. You state one measurable goal — “improve inference speed”, “raise this benchmark” — and the lab turns it into a research program: a contract made of instructions the loop follows and a pinned environment it runs in, so every result is reproducible.
Then the loop turns — plan → design → gather → run → analyze → report — and each turn is a real experiment: a falsifiable hypothesis, run on your machine, on its own git branch, ending in a measured number and a verdict — supported or not supported. The honesty rule is structural: if nothing can be measured, the experiment reports that it cannot run rather than handing you a recommendation. Never a guess, never a canned result.
The loop keeps working on your goals while you're away, and everything it finds is written into the knowledge base — so the next experiment, the docent, and your chat all start from what the last one learned.
02 · Model building
From streamed teachers to models you own.
The lab's deep mode runs on QueueLLM, which streams a large teacher model off SSD a layer at a time instead of holding it in memory — frontier-scale reasoning on the machine you already own. The tuning loop uses the same experiment-per-git-branch ledger as autoresearch, so a model change is a measured claim, not a vibe.
And when a subject settles, Nucleus distills it toward a compact model that runs where you are. The pipeline is built end to end, and today it runs deliberately, by hand.
03 · Knowledge curation & chat
One knowledge tree, and everything feeds it.
The center of the blueprint is a single knowledge base: research findings, curated sources, and mounted Worlds — the gate-checked ground truth the agents develop on — all land in one tree that feeds the wiki, the notebooks, and every agent. A curator agent vets what comes in; nothing enters the tree unreviewed.
And you talk to it. Chat is grounded in the same tree, with Buddy — the lab's companion agent — alongside a small crew: a researcher, a curator, a browser. Ask what the lab knows, ask what it found last night, or just ask it to explain something at the depth you want.
The manual and the AI dictionary live in-product too — the lab documents itself.
Go deeper
This page is the short answer. The long ones: how the blueprint bolts together, the slide deck, and the essay this page distilled.