01 · The System

A self-hosted operations platform and ARIA, an operations layer: work comes in, gets processed by rules and AI, and consequential actions wait for a human yes.

  1. The System: A self-hosted operations platform and ARIA, an operations layer: work comes in, gets processed by rules and AI, and consequential actions wait for a human yes.
  2. Intake and Information: Alerts, schedules and requests arrive from the systems ARIA watches, each enriched with the context needed to decide.
  3. Processing: Rules handle the known cases; AI routes the ambiguous ones. Circuit breakers stop runaway behaviour before it spends money or breaks things.
  4. Human Approval: The heart of the design: anything consequential pauses here. Approve, edit or reject — the system waits.
  5. Action and Reporting: Approved work executes, results are verified, and everything lands in reports: health checks, costs, backups, incidents.

Mosaic · Empire OS · ARIA · 2025–Present · AI and Digital Systems

Mosaic — AI Operations Platform

Founder and Builder

Private Repository

Mosaic is the operations system I built so I can run my own ventures with leverage a solo founder does not normally have. The platform and its safeguards run today; the agent business logic is still being extended venture by venture. The first tenant is my own set of small businesses, not a paying customer base.

Mosaic operations pipeline — concept visualization
Concept visualization of the pipeline. Not a screenshot of the live system.

Mosaic, Empire OS and ARIA

Three names, three layers. Mosaic is the product — the terminal an owner-operator actually uses. Empire OS is the self-hosted platform underneath: the containers, data, governance and monitoring that make it safe to run. ARIA is the agent layer on top — the workers that do the content, research and operations tasks.

Context

Running multiple ventures needs automation, but automation that can spend money or change live systems has to be safe. The hard part is not calling a model — it is making sure consequential actions clear a policy check, a budget and a human yes before they run. Building that layer of safety was the hard part of the work.

What I built

  • A self-hosted, multi-tenant platform (Empire OS) on a single VPS, deployed with Docker Compose
  • A governance overlay: an action gateway, a policy engine and a model-routing proxy in front of every consequential action
  • A human-approval gate and a tamper-evident, hash-chained audit log
  • Per-tenant daily budgets so each venture has its own cost ceiling
  • An agent layer (ARIA) running a content, research and publishing pipeline across many platforms
  • Full observability — per-agent LLM tracing plus metrics, logs and alerts

How it works

Each agent action enters the gateway with its tenant, type, chosen model and estimated cost. The gateway runs it through policy, routes the model call through a proxy with caching, holds consequential actions for approval, executes the approved work, and writes a hash-chained, tamper-evident record.

Task intakeaction-gateway (FastAPI)
Policy checkOPA — Open Policy Agent
Model routingLiteLLM proxy + Valkey cache
Human approvalgate on consequential actions
Actionn8n workflows · tools
Audithash-chained audit_log

Cross-cutting: per-tenant daily budget (cost control) · Langfuse + Grafana/Loki/Prometheus (observability) · Falco (runtime security) · all self-hosted behind a Cloudflare Tunnel.

Technical architecture

Ingress & routing
Cloudflare Tunnel · Traefik · socket-proxy
Data
PostgreSQL 18 · Valkey (Redis) · PgBouncer
Orchestration
n8n · a scheduler and worker/agent poll loop
Governance
action-gateway (FastAPI) · OPA policy engine · LiteLLM model proxy
Observability
Langfuse (per-agent LLM tracing) · Grafana · Loki · Prometheus · cAdvisor
Safety
Falco runtime security · Alertmanager · hash-chained audit log

The agent pipeline

The content and revenue pipeline, from inputs to learning. Each box is a real agent in the system, and every consequential action is held for human approval before it runs.

Inputs

  • trend_miner
  • seo_research
  • lead_capture

Make

  • draft + edit pass
  • image_maker
  • video_maker

Distribute

  • social_publisher → 11 platforms
  • etsy_integration
  • email_marketer

Learn

  • learning_loop
  • audit_log

Real agents from the codebase. Private repository — shown as an architecture map, not a screenshot.

The operator terminal

mosaic — operator view · running since 2026-02 · policy checks and human approval ON

● runningtrend_minertenant: fav_designslast run ▓▓▓▓
● runningsocial_publisher11 platformsqueue ▓▓
● runningseo_researchtenant: aureellast run ▓▓▓▓
◐ awaiting approvalemail_marketeraction held at gatecost est. ▓▓▓
● runninglearning_loopall tenantsaudit row ▓▓▓▓▓ (hash-chained)

Reconstruction of the live terminal — real agents and tenants, values redacted. Not a screenshot.

What runs on it

Several systems that used to be presented separately are modules of this platform. They live here because that is what they are.

Content and publishing

Trend research feeds drafts, a person reviews them, and approved pieces go out to eleven platforms.

Career operations

Job aggregation and scoring, tailored documents per role, and a document pipeline that renders clean, ATS-readable output.

Client operations

The Aureel Gems engagement runs on this platform: listings, content and routine operations for a real client, with permission.

Support agent

A knowledge-grounded support chat concept, planned and scoped. Still a concept, and labeled as one.

Portfolio build system

The research, build and review tooling that produces this website and its checks.

Operations layer (ARIA)

Alerts, health checks, cost and backup monitoring, circuit breakers and reporting across everything above.

A real action, audited

The gateway accepts an action, checks the tenant budget and policy, and records it in a hash-chained log, so any alteration of the history is detectable. Sanitized, from the runbook:

POST /actions
{
  "tenant_id": "fav_designs",
  "action_type": "content_generation",
  "model": "haiku",
  "estimated_cost_usd": 0.01
}

# audit_log: id | table | operation | row_hash (chained)
# each row's hash includes the previous row's hash — tamper-evident

Private repository — code shown as a sanitized excerpt, no secrets, no live link.

Operating model

Right now this is a dogfood pilot: it runs my own small businesses, on my own VPS, with no paying tenants. The direction I am building toward is the same self-hosted, ownable stack in the hands of other owner-operators — indie founders, small acquirers, holdco-of-one operators who want real leverage without hiring a team — but that is the intended next step, not something I have sold yet. Where it builds product for a company, the model is equity in what gets built, not a subscription.

Current status

The platform and its governance layer have run continuously since February 2026. The content, revenue and intelligence pipeline executes from intake to publishing and has done real work for my own ventures: published content and live operations tasks. Tenants are dynamic: today the platform runs my own ventures, and it is being generalized so other owner-operators can run theirs. The agent logic is extended venture by venture.

Back to the Work