Jason Dicken
Case study 04 · Restaurant inventory / AI-assisted application developmentActive development

Cobblestone Galley

Inventory and ordering for the kitchen on a towboat: offline-first, shared on one tablet by the chef, the cook and the captain. It's also my clearest example of coordinating more than one AI coding agent on a single codebase.

The problem

Barge and towboat galleys order food by phone and paper. Nobody can see what's on the shelf, and the captain approves spending without a clear picture of what was ordered or why.

And the boat is often out of signal range, so any app that needs a connection to work is useless exactly when it's needed.

The solution

An offline-first app the galley crew shares on one tablet. The chef counts stock and requests what's below par, the captain approves, and the purchase order goes out by email when there's a connection. Inventory, PAR levels, purchasing and reporting live in one place and sync when the boat is back in range.

One requirement shaped the whole data model: galley vendors are shoppers, not warehouses. There are no line-item prices to import, so the app records spend as the typed receipt total instead of pretending to know unit costs.

My role

  • RequirementsDefined the chef, cook and captain roles, the approval chain and the offline behavior from how galleys actually order.
  • DesignSet the data model and the decisions that constrain it, recorded in a dated, locked decision log.
  • Agent directionRan two AI coding agents in parallel in one codebase, with scoped tasks and mandatory written handoffs.
  • TestingReview each handoff and test the workflows before the next task is assigned.
  • DeploymentAndroid builds through Capacitor. Not yet in a working galley.

How the two agents share a codebase: a written baton protocol, where only the current holder may write, and a decision log the agents read before working. A constraint like "no line prices anywhere" is the kind of thing a model can't infer from the code, so it gets written down once instead of repeated every session.

Technology used

  • React
  • TypeScript
  • Supabase
  • Postgres
  • Capacitor
  • Android
  • Offline-first sync
  • Claude Code

Key features

  • Inventory & PAR levelsCounts against par, so the order is what's actually missing.
  • Chef-to-captain approvalThe chef requests, the captain approves, and the purchase order goes out by email.
  • Offline-firstWorks with no signal and syncs when the boat is back in range.
  • One shared tabletBuilt around the crew sharing a single galley device.
  • Receipt-based spendRecords spend as the receipt total, matching how galley shopping really works.
  • ReportingSpend and ordering history in one place instead of on paper.

Application screenshots

Screenshots from the current build are being prepared. This page will show real interfaces only, never mock-ups.

Demo

A recorded walkthrough of the chef-to-captain ordering flow will be posted here.

Development status

  • Working prototype
  • Active development
  • Internal production
  • Customer-deployed

In active development. Android builds exist; it hasn't shipped and isn't in use on a boat yet.

Results

No operational results yet, because it hasn't shipped. What it has produced is a working method: two coding agents building one app in parallel without overwriting each other, held together by a baton, scoped commits and a decision log any new session can read.