Munim.
AI back office for restaurant groups · a Manager + 8 specialists · nothing changes the books without approval
“Restaurant owners don't need another dashboard. They need someone to tell them where the money went — and never touch the books without asking.”
Small food businesses lose money in places nobody looks: a supplier's price creeping up, waste spiking on a Tuesday, a delivery-app payout that doesn't reconcile. An LLM can spot those patterns, but an LLM that can write to the books — or that invents a number — is worse than no help at all.
Split the system by trust. Code owns every write: ledgers are append-only, mistakes are corrected by a reversal with a reason, and a database trigger refuses any change without a named approver. The model only reads snapshots and proposes. Its prose may not contain digits that don't point at a snapshot cell, so every figure in an answer is a reference you can tap open. If a number can't be traced, Munim says "cannot verify" instead of guessing.
The choices that mattered, with the reasoning at the time.
- The AI has no write path
No database client, no write grant on the book tables, and a trigger that rejects unapproved writes. Safety comes from the architecture, not from a prompt.
decision · 01 - Numbers are references, not text
Model output can't contain a free-floating digit; figures are filled in from snapshot rows. Hallucinated arithmetic becomes structurally impossible rather than unlikely.
decision · 02 - Country packs instead of branches
Currency, tax lines, receipts and compliance calendars come from an India or Australia pack, so the same code runs both markets without a single if-country check.
decision · 03
Submitted to the Google Cloud AI Builder Cup 2026 (result pending) and, as Munim Business Graph, to FalkorDB Graph Hacks — where the books are projected into a per-business graph, rebuilt in a staging graph and swapped in atomically, and the AI reads only through bounded read-only Cypher templates.
- specialists
- 8 + a Manager
- AI write grants on the books
- 0
- tests
- ~2,600
- country packs
- India · Australia
Next I'd put a real shop on it. Everything is honest about being synthetic today, and the most useful test is one owner using it for a month.
- Next.js apps: storefront, staff console, developer portal
- Python intel service: Manager + 8 specialists on Google ADK + Gemini
- PostgreSQL books: append-only, approval-gated by trigger
- FalkorDB supply-chain graph, read-only to the AI
- Cloud Run (asia-south1), Cloud Build, Secret Manager