AI that admits what it doesn't know.
Univault Technologies is an AI research company in Salt Lake City, Utah. We have spent years on the failure that matters most in real work: AI that is wrong with confidence. What came out of that work is infrastructure that asks instead of guesses — and we help teams put it to work where "probably" isn't good enough: finance, hiring, operations.
Three releases, one principle
Everything we ship carries the same discipline: when the system is not sure, it says so — to a person, before anyone relies on the answer.
Inference Gateway
A production gateway for AI agent workloads, operated under live traffic for a paying customer.
When it is not sure enough to stand behind an answer, it refuses and says so explicitly — a signal an integrating team cannot silently ignore.
In production
Expenses
Say one sentence about a purchase. It comes back as a determined, filed, defensible record, checked against your own policy.
When it is not sure which clause applies, it asks a person instead of guessing. A finance record that says "probably" is not a record.
Early access, by invitation
Roles
Hiring claims checked against evidence, clause by clause, before anyone acts on them.
"Cannot tell" is a first-class answer. The system would rather admit silence in the evidence than manufacture a judgment about a person.
Early access, by invitation
How we work: help first
Before anything is sold, we sit with your team and map where this class of AI genuinely holds up in your work — and where it fails quietly. You leave with that map either way. We take on a small number of these engagements at a time.
If your team runs work where a confident wrong answer costs real money — or real trust — we should talk.
Talk to us
Now open to selected partners
The gateway has a name: Bees
We have run Bees for a while — behind our own releases, and behind our first paying customer. It runs the routine parts of an agent's work on a hive of small models, and calls in the frontier only when the work demands it. For teams whose automation is hungry for tokens, that is real money on the line — and every request is metered, so the savings are yours to verify, not ours to claim.
The field is arriving at the same conclusion from the opposite direction: specialists beat one generalist. The difference is where the specialists live. Others keep theirs inside one building. We think they belong spread out — because the future of automation is not just cheaper tokens. It is accountability: AI that knows its limits, keeps an auditable record of every request, and brings a person in while the automation keeps running.
The research underneath
Measurement research backs every release above. A few results, each one audited against our own evaluation discipline before it was allowed on this page:
26.86 µs
Real-time reflex perception
Constant-time perception query at p50, measured over 10,000 queries on NVIDIA Jetson embedded hardware, in a fixed-size store that does not grow with what it learns.
0 bits changed
Certification-preserving field learning
The SHA-256 digest of a certified 941,316-parameter model, unchanged after 1,000 field enrollments while the learned pattern store changed at every one — verifiable by the operator with standard tooling.
42 / 42 checks
De-identification, independently re-verified
Encodings not invertible to the person: speaker identification at 0.00 percent against 1,306 speakers, genomic identification exactly at chance — re-verified end to end at a different seed.