while-ai designs and builds AI software: agents, document automation and private LLM systems. And we help teams figure out where AI really pays off.
From the first prototype to software that runs in production every day.
Clear answers on where AI pays off, and where it does not.
A working version now takes hours, not weeks. So we stop guessing up front and start experimenting.
Plan for months. Build once. Hope it fits.
Build in hours. Try several. Keep what works.
You bring an idea, or we find one together.
ExperimentQuick versions side by side, tried on real data.
BuildOnly what proves itself lands in the version that always works.
↺ IterateEvery day, back to the next idea.
An idea is cheap to try and expensive to build wrong. So before anything is written we make sure we understand it: we think it through, ask instead of assume, and reduce it to parts we know.
Every change starts as a written brainstorm: the problem, the options, the open questions. An unclear requirement is asked, never assumed.
If a library, a tool or an open-source project already solves it better than we can, we use it. We build only what nobody offers.
We break a problem down until its parts are standard components with proven theory and mature implementations: a parser, a state machine, a graph search. Then we assemble and test instead of invent.
A working version now takes hours, not weeks. So we build quick variants side by side, run them on your real data and let the results decide.
Every build has a corner for spikes and prototypes. They run on agent-manufactured mockups of your actual documents, logs and numbers, so nothing sensitive leaves your house.
What survives the experiments is written down as a one-page spec: the user story, the parts, the objects behind them. A machine checks it: the story parses, every object exists in the code.
Before we commit to a design we play its next three growth steps: a new format, a new view, a new rule. A healthy design multiplies; one that swells gets cut before it is built.
Agents write most of the code. They work under rules: in isolated copies, in small steps, with tests, and nothing lands without a review.
Our process engine holds every document and every sprint, runs the doors that create, validate and close them, and witnesses every change on disk. An agent takes a task to complete; a human holds the locks.
A change is done when its tests pass: end to end, on real inputs, no mocks. Trunk-based, in small steps, so the main line always holds a working version.
Every result is reviewed by a second model before it merges, then by us. You choose the conduct: approve every phase, or let the sprint run to its holding point and review the whole.
Shipping is not the end of the loop. Everything we learn is written down where the next sprint will find it, and the version that always works stays that way.
Closed work goes to the changelog, decisions to the decision log, open items to the issue list, automation ideas to the improvement list. Nothing lives only in someone's head.
We never delete: files, code and drafts move to an archive. Secrets never enter a repository. Parked topics wait in an icebox and are read before anything new is started.
A spec is not a document that goes stale: it is what the code must be, and every change starts there. Checked against the code on every sprint, it keeps the architecture consistent and scaling as designed.
Self-hosted and open source wherever possible, no vendor sign-ups. Tested, documented and handed over, so you can run it, read it and change it without us.
while-ai is built the way we think software companies will work: a small core team, a fleet of AI agents, and clear rules and reviews that keep them in line. For you that means the speed of a startup, the output of a large team, and you always talk to the people accountable for the work.
We show you something that runs, as early as possible.
Self-hosted and private model options whenever confidentiality matters.
Tested, documented, every decision on record, and handed over, so you are never locked in.
Tell us what you are working on. A short email is enough to get started.