What this is actually for
The fear was never that it wouldn't work. It's that one day it says something you would never say, to a customer, in your name.
What we built for ourselves
Every piece of our business data carries a sensitivity tier, and the tier travels with it into every generated message.
A hard approval gate physically blocks sends. Not a policy — a gate.
Approval authorises one specific batch, never a pattern. That's the failure mode that produces forty unreviewed messages.
One word stops every downstream automation in the chain and reports back what it stopped.
What you get
- —Data classification tiers, set by you and encoded — not inferred from a vendor default
- —A hard approval gate on outbound, with a named approver list
- —Single-use send authorisation and a one-word global stop
- —A full audit trail of what was generated, approved, sent and blocked
- —Classification running on local hardware, so message content never leaves your premises
Hear how it got built
We recorded the build of each of these, including the parts that didn't work.
- Episode 2 — Ask The Other Onerecording
- Episode 4 — The Brain Maprecording
Episodes publish as they're cut. Nothing here is a teaser for something that doesn't exist — the systems are already running.
What this doesn't do
The gate does not make the model correct. It makes mistakes visible before they leave the building — it turns an incident into an annoyance. Anyone selling you a system that can't be wrong is selling you something that doesn't exist.
What it runs on
Built on NVIDIA hardware and software we own and operate, not a reseller relationship.
Where it's going next: Unbundling it, so it can be installed on AI you already have running ungated.
A member of the NVIDIA Inception program
Fifteen minutes, no deck.
Bring one specific problem in one sentence. We'll tell you whether it's worth building, roughly what it takes, and sometimes that the answer is don't build it.
Start hereTwo steps before we talk — it takes about five minutes.