Council
AI bounded to your organization. Across every surface.
It is as if all your tools just learned to talk to each other very carefully — with the creativity of your marketing team, the discipline of your grant writers, and the carefulness of your CFO.
Council is an AI-native system that works in service of a nonprofit leader. It is bounded by what your organization actually knows — your sources, your figures, your approved language, your prior decisions — and it brings that knowledge into every surface where you carry the institution: writing, sourcing, reviewing, approving, archiving, and noticing when artifacts have drifted apart.
You start a board memo, and Council composes a first draft from your figures. The reviewer you choose sees it in their queue. The memo holds until its source packet is attached. The ledger remembers every claim with its evidence and how recent it is. When the same claim shows up in next quarter’s funder report, Council surfaces whether the number has moved — for you to decide whether to update.
The work moves at the speed it needs to. The institution stays coherent without anyone holding it all in their head. The hours go back to the work only you can do — the conversations, the decisions, the judgment your role asks for.
Underneath
Underneath the surfaces, one institution.
Council is composed in three layers that share a single underlying model of your organization.
Memory.
Your sources, your figures, your approved language, your prior decisions, your commitments. Council holds these as structured knowledge that can be queried, refreshed, and reused — not buried in folders, not scattered across inboxes.
Composition.
Drafting, sourcing, reviewing, approving. Every artifact composes from the memory layer with sources attached, voice intact, and audience-appropriate attribution. The work moves through your organization’s existing approval shape, recorded as it happens.
Coherence.
Across artifacts, over time. Council notices when this week’s draft contradicts last quarter’s report, when retired language has crept back in, when a commitment is past due with no follow-through. The system watches the system.
All three layers run on an AI-native kernel built for institutional work. Every event — every draft, every approval, every claim refresh — is recorded as part of the institution’s working history.
Built today
120 typed kernel object families model the institution.
Sources, claims, decisions, commitments, approvals, audit events — each kind of institutional thing has its own structured place that holds together as the work continues.
129 event types record what happens to it.
Every change — a draft saved, a source refreshed, a phrase retired, an approval granted — is recorded as a discrete event you can look back at.
8 AI route endpoints. 47 named institutional skills. Every model call bounded to your workspace.
Council never reaches outside your organization for context. Each AI call runs against your own sources, figures, approved language, and past decisions — shaping the draft through one of 47 named moves (sharpen, soften an ask, lead with the numbers, center the population served). Each move is versioned like software, and its output quality is observed over time.
155 automated checks run before every release.
Before any change ships, 155 checks watch for regressions, drift, and breakage across the whole system.
Why this approach
If Council works at scale.
AI is most useful when it is bounded. The same model that hallucinates against the open web becomes reliable when its context is one organization’s actual sources, figures, and prior work. Council is the system that holds that context coherently across every surface a nonprofit operates — and gives every model call the institutional ground it needs.
The nonprofit sector is a good place to build this first. Nonprofits write constantly — board updates, funder reports, public statements, internal briefs — under conditions that are unusually unforgiving. Outcomes are scrutinized; commitments are remembered; the same claim travels across audiences in different appropriate framings. Staff capacity is thin. The cost of getting the institution’s own knowledge wrong is paid in board meetings, funder calls, and lost weeks.
The pattern generalizes. Any institution that operates through written artifacts — health systems, universities, agencies, advocacy organizations, mid-market companies with a compliance footprint — needs the same kind of layer for the same reason. Council is being built for nonprofits because that is the most useful place to make it real.
For investors and operators looking at where AI-native systems take root next: the answer increasingly sits inside the context of one organization, where the model has somewhere to stand. Council is betting it has found the right shape that generalizes with scale.
Contact
Get in touch.
I am Ravi Narain. I conceived, designed, and am building Council — currently solo in Seattle. If this interests you — as a nonprofit leader who recognises the problem, as an ML researcher with a better thesis than mine, as an investor looking at where AI-native systems take root in organizations, or as someone building in adjacent territory — write directly. The reply will not be quick, but it will be a real one.
Council is in early development. Access will be invite-bound when the first reviewers are ready; there is no public sign-up yet. The conversation can start now.