Work with us.

A new architecture for language systems. Today's frontier models entangle the world, language, and reasoning into a single weight space and reconstruct everything on every query. Calliope distributes cognition across regions modeled on the human brain, so knowledge lives outside the weights, reasoning is a separate step, and the system can be inspected at each boundary. The result is a system that grows without retraining, costs less to run on routine work, and can tell you when it doesn't know.

Through early 2026 the market established that general-purpose agents at scale can cost more than the people they were meant to replace, and that the bill grows even as per-token prices fall. The cost structure of the monolith is the problem. We're not building a cheaper version of the current thing — we're building a different cost structure for intelligence, which is the kind of shift that creates a market rather than competing inside one.

The fuller market analysis is on our research page.

Ioma Labs was founded in 2026, formalizing work the founders had been building toward for about a year prior. Calliope exists as a working prototype in closed testing, and we benchmark it against frontier single-pass models. We're candid about where the approach wins today and where it doesn't yet — that honesty is part of how we work.

Our investors — benefactors, in the language of Project Ancilla — get early access to the work, regular progress against published benchmarks, and a stake in the architecture as it scales from prototype to product. Specifics on round, terms, and use of funds are shared directly. Email us and we'll set up a call and send the deck.

The architecture is the moat, and we treat it that way. Patent work is underway, which is part of why our public material describes what the system does rather than how it does it. Under NDA we can go deeper.

Because cognition is distributed across separate regions, the system produces an inspectable record of what it understood before it says anything — you can audit how a conclusion was reached, not just read the output. It computes whether it knows something apart from what the answer is, so uncertainty is an honest signal rather than confident invention. These are first-class properties of the design, not bolted-on safeguards.

The common-case path runs on a few small specialists, with the heavy reasoning step recruited only when a question genuinely needs it. That makes a much smaller compute footprint viable than a monolithic model requires — relevant wherever data-center-scale inference isn't an option. We can discuss specific deployment envelopes directly.

Knowledge is held outside the weights and retrieved, rather than baked into one shared space, so what the system can draw on is governed rather than diffuse. Control is structural — specialists know only their own job — rather than a prompt instruction that can be talked around. Specifics are a conversation for a cleared setting.

We're an early-stage research company and we treat government and defense work accordingly — scoped, deliberate, and under the appropriate agreements. Michael Musilli leads contracting and legal on our side, including government acquisition pathways. The right first step is an email introducing your office and interest, and we'll route from there.