ABOS — a sovereign business operating system

I design and build ABOS: a single AI “Mind” that understands a business, acts through tools, notices what changes, and learns from the outcome — running end to end on the client’s own hardware, on local models. The point is what it doesn’t do: no data leaves the network, and no outside AI vendor sits in the loop. The human stays the decider.

I built the whole substrate myself — no agent framework, no off-the-shelf memory layer — on Postgres and self-hosted models.

The core loop

A structured model of the client’s business; an observation engine that watches it continuously; a proposal ledger so every action is tracked, not silent; an operating cadence with an attention budget so it acts when it matters; outcome learning that adjusts to what actually happened; and a trust-and-risk gate with tiers and rollback, so autonomy is earned rather than assumed.

The reasoning core

Mind and its per-turn loop, on local models, wrapped in a plan → judge → push-back → learn cycle so it argues with itself before acting.

A memory system, from scratch

Rather than bolting memory onto a black-box model, I made retrieval smart over data owned in typed Postgres. It behaves like human memory: important facts anchor the system, attention pulls in deeper context, and an association graph links related records nobody connected by hand. Retrieval ranks on a distribution of weighted signals — semantic similarity, importance, association strength, and recency — instead of flat nearest-neighbor search, and fuses keyword with vector search so exact names and codes match as well as meaning does.

Self-hosted model serving

I run the models, not an API — multiple instances load-balanced across GPUs with zero-downtime restarts, and a voice design where one model plays two roles in parallel instead of paying for two.

A system that extends itself

A loop that detects a need, researches it, plans, builds, and self-tests, escalating to a human at the edges.

Business logic as configuration, not code

Workflows and rules live as data loaded into Mind’s context, so a new client is configured, not rebuilt.

The plumbing

An event bus and tool layer; an LLM router that handles model choice, failover, security tiers, and cost in one place; a tool registry where a new capability is a decorated function; chat, voice, and email channels; a client portal; and security from day one — encrypted secrets, multi-tenancy, audit trail.

The systems depth of twenty years in infrastructure, pointed at applied AI — for teams that need it to run on hardware they control.