OnnAI offers a toolkit, methodology, and specialized AIs for automated iterative refinement. Its ecosystem of 30 composable tools lets you build custom refinement loops — a stochastic search across solution spaces powered by non-deterministic software. You define the standard. AI does the work. You can see how it got there.
OnnAI teaches organizations how to use AI to build iterative refinement into their workflows — then gives them the tools to do it.
Half-day and full-day sessions that teach the kaizen methodology. How to write personas, define loss functions, build loops, and review logs. For teams of 5 to 500. Remote or on-site.
We build your first kaizen loops with you. We audit your document workflows, write the personas, configure the tools, and run pilot programs until your team owns the process.
The full toolkit — 30 tools that compose into custom kaizen loops for your domain. One license, your infrastructure, your data. macOS and Linux.
Each Onnai tool does one thing. Composed through pipes. You edit plain text files, in English. No Python, no code.
Composable AI. Cloud models, local models, and Onnai's specialized models, in any combination.
Memory for AI. Intent-first, moments instead of commits. Why instead of what. Full tree snapshots, content-addressable. Undo, branch, search. So easy, even an AI can use it.
A shell rebuilt for AI. No quoting nightmares, no $ expansion. times runs N iterations. while runs until completion. each maps over file sets. Built in RPN calculator.
Optimized messaging for kaizen coordination. Hash-chained audit log. Sockets locally, multicast on LAN, encrypted over the Internet. Build bots in one line.
Every kaizen loop is assembled from composable tools. Different domains, different pieces, different loops. The building is part of the engagement — we help you design the loops that fit your workflow.
The role, the standard, the review criteria. This is the loss function — it defines what "better" means for this document type.
You are a regulatory affairs specialist.
Review against ICH E6(R3). One change
per round. Read .context. DONE when
submission-ready.
Which commands onn can run. Minimal surface area. Enforced by the OS-level sandbox — onn proposes, the sandbox disposes.
ito
cat
grep
wc
squawk
The optimization target for this run. Specific criteria that tell onn what dimension to improve along.
$ 10 times onn protocol.md
"improve — ICH E6(R3), completeness,
no ambiguity in endpoints"
Every change logged with intent. Content-addressed. Immutable. The audit logs, review record, and revision history — all in one.
$ ito history
a3f912c define primary endpoint
b7e44d1 add stopping rules
c912fa3 clarify inclusion criteria
DONE after 3 rounds
Onn talks to any model — our specialist AIs, cloud APIs, local inference, or any combination. We offer unified billing for organizations that want one vendor. For sensitive documents, run everything locally. Your data never leaves your machines.
OnnAI's pre-built specialists you call by name. Editor, Writer, Auditor, Translator — an ever-growing library. Some are fine-tuned models, some are proprietary system prompts. No persona files to write. Billed as inference.
One invoice for all inference across your organization, or bring your own API keys. No per-seat LLM subscriptions, no surprise bills.
For regulated industries, classified environments, or anywhere data can't leave the building. Onn connects to local models — Llama, Mistral, Qwen, whatever you run. Same kaizen loops, zero network calls. We help you set it up.
Travel billed separately for on-site events. Volume licensing available.
Unified inference billing available as an add-on.
The same pattern applied to every domain where text artifacts are iteratively refined against a definition of "better."
Enter your email to access the deck.