AI will increasingly help humanity run a lot of our world.
Our projects focus on understanding and improving how people can work with and manage AI to run the world better.
We publish meaningful parts of our work on our blog and soon on GitHub as open or shared source.
PILOT · PROTOTYPE
Keep AI work
on course.
Air traffic control for your AI work.
See what needs you. Recover the context. Correct the course.
Explore the examplesCatch the mistake early.
Prepare tomorrow’s client review
Drafting the meeting brief
Present last quarter’s forecast.
What is it assuming?
That forecast is still current.
Finance approved a revised forecast yesterday.
Finance update · approved yesterdayPick up the work without starting over.
One job with a clear result: a meeting brief using the latest approved forecast. Another agent can check open customer issues in parallel.
What can the current prototype capture?
Coverage depends on the connected tool. The current adapter keeps task fields and file references, but does not capture a complete transcript or a stable ID for each attempt. Referenced files are not verified. The client review above illustrates the concepts; it is not a connected account.
How memory, agency and judgment help with the client review
Memory
“Finance approved new numbers.”
Agency
“The draft uses the old forecast.”
Judgment
“Use the approved update.”
Check the numbers before presenting them.
The next brief starts with the corrected forecast.
Tell the team once.
Each role knows what changes.
Which invoice does this settle?
Match it to invoice 1042.Should I chase the customer?
No. Cancel the reminder.Is the payment still overdue?
No. Update the cash report.Keep corrections useful across tools.
Web context revisions
Mac briefings
Live-account validation
DGX Spark: inference + video
