Domain AI
Domain AI is what Solidarity Intelligence actually is, once it stops being an argument and becomes something running: not one large, general-purpose model meant to serve everyone everywhere, but a set of separate, smaller intelligences, each built to fit one field of real work — a local food store, a citizen solar programme, a community currency, community organising.
Not a giant model, but an intelligence that fits
The movement’s own framing states the choice plainly: build “domain AI, built by our own hands — not outsourced to Big Tech. In the spirit of appropriate technology and DIY: not a giant model, but an intelligence that fits us.”
Two ideas are doing the work in that sentence. Appropriate technology is the older idea — associated with development economics and grassroots technology movements — that the right tool for a community is the one matched to its actual scale, resources, and skills, not the most powerful or most prestigious tool available. DIY (do-it-yourself) is the practical stance that follows from it: rather than waiting for a vendor to build the right tool, or accepting whatever a large platform decides to offer, the community builds and modifies its own.
Applied to AI, this means resisting the assumption that bigger and more general is automatically better. A single food co-operative does not need — and cannot afford to run — a model built to answer every question in every domain for millions of users. It needs something that knows its own products, its own members, its own store operations, and can act on that knowledge directly.
The three-layer approach to sovereign AI
The movement’s stated approach to building sovereign AI combines three layers rather than attempting to build a large language model from nothing:
- “The sword we use now” — commercial large language models, used as they exist today (Gemini, Claude, and others), because they are already the most capable tools available and refusing to use them would only slow the work down.
- “The sword we are forging” — open-source language models that can be trained and fine-tuned over time, gradually reducing dependence on any single commercial vendor.
- “The irreplaceable core” — on-the-ground data, automation, and tacit knowledge: the actual sales records, member relationships, store procedures, and accumulated know-how that no outside model, however large, has access to.
The explicit point of this framing is that sovereignty does not come from trying to build a model from scratch — a 32-billion-parameter model, for instance, is already well beyond what a small co-operative can train outright — but from connecting the best available large language models to a community’s own on-the-ground assets. The model is rented where renting makes sense; the ground truth stays owned.
Working examples
This is not only a design principle; it is implemented in several separate systems, each scoped to one field rather than built as one general assistant:
- Poomai, grown up around the Jijok and Gwanjeo local food stores in Daejeon — see Data Sovereignty in Practice for its technical structure. It does not merely answer questions; it sends real customer text messages and prints real price tags, described internally as “physical AI on the web.”
- A citizen renewable-energy AI, supporting automation work for citizen solar generation.
- Poome, a solar-sharing AI serving that same energy work through a separate chat interface.
- CoAI, an AI built for community organising work rather than for retail or energy.
Each of these is a distinct system serving a distinct field, rather than modules bolted onto one universal platform. That separation is itself part of the “appropriate technology” stance: a tool sized to the problem in front of it.
A companion, not a replacement
Domain AI in this movement is explicitly framed as companion AI — “a tool like a bicycle, extending human judgement rather than replacing it.” The aim stated for these systems is not autonomous decision-making but augmentation: work that a person could in principle do, done faster or more reliably, while the person stays in charge of judgement calls.
Who can build it
Because the aim is DIY rather than vendor-dependent, the movement treats the ability to build domain AI as something that should be teachable, not restricted to specialists. Materials are gathered — from concepts through to hands-on practice — through the AI Organisers’ Commons, explicitly so that “an organiser who has never written a line of code can build AI with their own hands.” Domain AI, in other words, is meant to be something a food co-operative, an energy co-operative, or an education programme can build and maintain itself, rather than something it must always purchase.
Relationship to data sovereignty
Domain AI and data sovereignty depend on each other. Building a domain-specific system only produces real sovereignty if the data behind it — sales records, member relationships, Q&A memory — stays under the community’s own control rather than being absorbed into a vendor’s platform as the price of using the tool. The three-layer approach above exists precisely to keep that separation clear: rent the model, keep the ground.
Korean-language sources
- https://poomasi.org/ — home site (Korean)
- https://poomasi.org/en/ — English landing page, “What we are trying to build”
- https://poomasi.org/llms.txt — identity document (Korean)
- https://activist.poomasi.org — AI Organisers’ Commons
- https://coai.poomasi.org — CoAI
- https://chat.solarshare.kr — Poome, solar-sharing AI
- https://energy.poomasi.org — citizen renewable-energy AI