A business is a conversation
its own systems never have.
Every growing business runs on disconnected islands. Each one is fine on its own. Together they hide the answers that matter.
The loop: read → model → reason → act
Four stages, running continuously. This is the same closed loop documented at Palantir, Glean and JPMorgan — shrunk to fit a real SMB.
READ
Collectors sync every source on a schedule. No retyping, no exports.
invoice tracker → invoices
client files → organisations
MODEL
Facts become connected objects: the digital twin. Links are the magic.
meeting ↦ organisation
follow-up ↦ engagement
REASON
Claude reads the twin through MCP — every query filtered by who's asking.
we last discuss?" → one
answer, joined live
ACT
The AI proposes. A human approves. The action writes back. Audited, always.
→ follow_ups + audit_log
Eight layers. Each one boring.
Together, a brain.
The verified reference architecture — every layer annotated with who provably runs it at enterprise scale, and what the twin uses at SMB scale.
ABusiness Systems
Where work already happens. Untouched — the twin reads them, it doesn't replace them.
M365 · WhatsApp
BIntegration
Connectors read from — and write back to — every system. MCP is the standard plug (Linux Foundation, adopted by OpenAI, Google, Microsoft).
as: Glean's 100+ connectors
CData Foundation
One queryable store for everything the collectors bring home.
as: Palantir Foundry datasets
DKnowledge — the Digital Twin
Objects + links + permissions. Customer ↔ Job ↔ Invoice ↔ Meeting. The ontology IS the product.
as: Palantir Ontology · Glean graph
EReasoning
LLMs retrieve from the twin (semantic + relational) and think with full context.
as: JPMorgan LLM Suite
FOrchestration
Schedules, triggers, and the human approval gates that keep autonomy earned, not assumed.
as: Klarna's agent on LangGraph
GExperience
Where humans meet the brain: chat, a morning brief, a dashboard. No new app to learn.
briefs/YYYY-MM-DD.md
HGovernance
Every question and action logged. Roles enforced by the database itself. Trust you can show, not claim.
as: Palantir AIP Evals
Same twin. Same question.
Different clearance.
Permissions live in the database, not in a prompt. Flip the role and watch what the AI is even able to see. Try it:
| Organisation | Engagement | Value | Last meeting | Invoice | Status |
|---|---|---|---|---|---|
| Summit Build Group | Stage 2 — Extension works | $18,400 | Variation approved · Tue | OVERDUE 12d | |
| Coastline Electrical | Fit-out — 3 sites | $7,250 | Site walkthrough · Thu | SENT | |
| Harbourview Cafe | Refurb — final stage | $4,900 | Handover call · last wk | PAID |
// Postgres itself refused. Not the app. Not the prompt. The database.
Why this matters: even a prompt-injected AI can't leak rows the database won't serve. The viewer still sees organisations, engagements and meetings — the business keeps moving — but money is structurally invisible. Column-level too: staff see an engagement's name and stage, never its value.
We run one ourselves —
live, on LUNARC's own business.
We eat our own cooking first. LUNARC AI is tenant #1; every client install is a re-instantiation of this pipeline with a different ontology. And the ontology grows with the business — our latest extension (a complete program-roster of 15 people, meetings auto-linked) was designed, reviewed and shipped in one afternoon.
Sources
client files
invoices · pipeline
Fireflies API
Twin DB
Supabase Sydney
8 tables + RLS
schema committed
Seed + Collect
16 meetings · 2 invoices
15 docs embedded
idempotent, proven
MCP Server
7 tools · 2 roles
audit on every call
✓ Connected
Daily Brief
Haiku, ~½¢/run
facts-only rules
first brief 13 Jul
Live walkthrough
ask for the 15 minutes —
see it running on
a real business
Value proven at every phase —
or the next phase doesn't start.
Most AI projects fail because they're built hype-first. A twin installs value-first: each phase must show measurable results before more is built. That discipline is the product.
Audit & ontology sketch WEEK 1
We sit with each department and map your world: what are the things, what do you track about them, how do they connect, who may see what. No software yet — just the questions that decide everything.
A one-page map of your business as objects + links, signed off by youFoundation WEEKS 2–3
Your twin stands up in YOUR own cloud account (Australian region) — you own the data and the keys, permanently. Collectors connect your existing systems. Nothing gets replaced.
Your data flowing in daily, permission walls tested liveThe read-only brain WEEKS 4–6
Ask your business anything in plain English. A morning brief per director. No automation yet — first, the connected view nobody in your company has ever had.
Your team answers real questions from the twin, unaidedGated actions FROM WEEK 6
One process at a time: the twin proposes, a human approves, the action executes and writes back. Follow-ups, chasers, scheduling — each earns its place.
Measurable hours returned per week, on paperEarned autonomy ONGOING
Where an agent's accuracy is proven over time, gates loosen — deliberately, one task at a time, with the audit trail as referee. The twin grows as your business does.
The pattern, proven by people
with more money than us.
US commercial growth YoY — driven by AIP, the ontology "brain" product.
Q3 2025 filingsARR reading whole companies into one permissions-aware graph.
TechCrunch, May 2026staff on LLM Suite; ~$2B reported annual savings, 3–6 hrs/wk each.
CNBC · Forbes (press)of workload handled by one assistant; resolution 11 min → under 2.
OpenAI case study"Company brain" for F&B — same loop, same LangGraph, same bespoke installs. Proof at start-up scale.
meuze.ai (stack self-disclosed)got nothing from bolt-on AI. Embedded, foundations-first firms: 3× likelier to see returns. That's the whole thesis.
29th Global CEO Survey, 2026Eight terms, no mystery left.
The list of things your business is made of (customers, jobs, invoices) and how they link. Designing it is the consulting skill; tables + foreign keys is all it is.
A live copy of your business's facts, structured so a machine can walk the relationships the way you do in your head.
Model Context Protocol — the USB-C of AI. A standard plug that lets any AI client use your twin's tools. Linux Foundation standard; OpenAI, Google, Microsoft all adopted it.
Text turned into number-coordinates so "similar meaning" is findable. Powers "where did the client mention the variation?" without exact words. Pennies per document.
Row-Level Security — the database itself decides who sees which rows and columns. Un-promptable, un-jailbreakable, because it never reaches the AI.
Safe to run twice. Collectors upsert on natural keys (a meeting's transcript ID), so re-runs can never duplicate data.
The AI proposes, a person approves, only then does anything execute — and the audit log remembers forever. Autonomy is earned per-task, later, with eval scores.
A phase-exit test that must pass before more money or ambition flows in. Our immune system against becoming a Gartner statistic.
See yours in 15 minutes.
Built by Tika Rijal — 15 years inside Melbourne supply chains.
Book a time with Tika →