What we build

Systems that cite their sources — or refuse.

We don't plug in someone else's API and call it a product. We build the model, the GPU code underneath it, and the guardrails that make it safe to ship.

Talk to us See the research →

The product line, in one breath

Two things you can buy right now (the readiness audits, FBL). Two we'll build with you (the no-bluff engine, agent fleets). Two we're honestly still researching (NovaTerra, the Friend).

Under the hood, all of it runs on QUANTA — the language we built for AI. Every instruction carries its real, measured price and can't crash or bluff: the AI writes the what; the compiler guarantees the how. See the research →

Get today

AI Readiness & Audits

Shipped & measured

Start where you are: a free readiness scorecard, then a strategy audit, then a compliance-grade AI audit for regulated firms. Honest about what AI is ready for today — and honest that only the first of the three has ever been used.

Free AI Readiness Scorecard — five minutes, three concrete next steps. Live and self-serve today. 🟢
Strategy audit — where AI pays off for your business, and where it won't. Offered and scoped; no client engagement delivered yet. 🔭
Compliance-grade audit — documentation an assessor expects, for safety-critical work. We hold no certification ourselves and have not yet delivered one. 🔭

Take the readiness quiz

Get today

FBL — idea → working code

Shipped & measured

A wiring spec compiled into a working application. Seventeen framework emitters are built; nine of them are proven byte-identical under 46 byte-identity tests. The same intent, many targets — and we never let the build-target count and the proven count share a number.

emitters built / proven byte-identical: 17 / 9
byte-identity tests: 46

Talk to us

Build with us

No-Bluff Knowledge Engine

Prototype

AI that answers from your documents with exact source citations — or refuses. No vector database, fully deterministic. The trust wedge for regulated work.

Cite-or-refuse: every answer is grounded in a source, or it declines.
Deterministic retrieval — same question, same evidence, every time.
Cite-or-refuse is enforced at the type level. It is a property under test, not a guarantee: the first test built from outside our own question sets found two confident lies, and we publish them.

Talk to us about a build

Build with us

Agentic Systems

Prototype

Director→Manager→Worker→Tool fleets driven by a deterministic planner that cannot fabricate an agent id — with a STOP button shipped before the fleet. Virtual offices that mint real artifacts.

execution-proven agents: 67
deterministic planner: 1
shipped before the fleet: STOP

See the research

On the horizon

NovaTerra — the world

Research

A typed-graph world you own, where your AI Friend runs on your own machine. Our flagship research bet — presented vision-forward and honest, never shipping today.

Read the research

On the horizon

The Friend / Friend Machine

Upcoming

A local, private personal AI — and a hardware appliance for people without a GPU. Local-first by design: your data, your silicon, your Friend.

Read the research