Integrallis

Independent AI R&D · Est. 2002

You own it. We proved it.

Integrallis is an independent research and development lab. We build AI infrastructure that runs on hardware you control — in your process, on your storage, on your device — and we publish the measurements that say when that is enough, and when it is not.

The research program Open source on GitHub ↗
Six
studies across four research programs
Nine
open-source projects, Apache 2.0
Two
products carrying the research
2002
independent since

Position

What we commit to, and where we say no.

The frontier is not the only axis of progress. The more consequential shift is intelligence getting small enough, cheap enough, and local enough to sit inside ordinary systems — next to the data, on the device, under the owner's control. Three convictions follow from that, and everything here is an attempt to test them.

  1. 01Ownership

    Everything we build runs on infrastructure you already control: inside your process, on your object storage, on your device. Retrieval, inference, and memory stay inside the boundary you already defend.

  2. 02Evidence

    Nothing ships on a model's confidence. Models are qualified against measured capability, code is scored deterministically, builds pass on real execution, and our papers publish their own noise floors.

  3. 03The boundary

    Local execution is not universal. Model size, memory, latency, and hardware decide where it stops working — and we would rather publish that line than pretend it isn't there.

Tracks of work

Five lines of inquiry.

01Applied AIPower-conscious inference, small and vertical models, mixture-of-experts optimization, and a qualification standard for routing between them.→ 02Software engineeringWhat context a coding agent actually needs, where self-testing quietly fails, and how agentic memory changes the work.→ 03Frontier questionsUniversal embedding formats, a separable reasoning core, shared experts across small models, and a common latent space.→ 04Human–agent interfacesAn operating system for agents, harnesses that restore comprehension to fast development, and learning paths built from what you save.→ 05Open foundationsVectors, models, model artifacts, and agentic memory — an owned AI runtime for the JVM, released under Apache 2.0.→

Research, productized

Where the work ships.

Launching shortly · moobi.dev

Moobi

A test-first app builder for Kotlin Multiplatform, and the first place our work on agent context, bounded generation, and measurable quality gates becomes something you can use. Opening to a first group of builders this week.

TDD by defaultExecution as ground truthAndroid + iOS
Product line