It’s Sunday, and Sarah Chen is doing math in her head.
She runs six restaurants. Right now three people are walking toward overtime, a Friday rush is about to be understaffed, an invoice sits uncoded, and a renovation is creeping over budget. All true this minute and she can see almost none of it. The data lives in five disconnected tools, and by the time it reaches a report, the week is gone. She’s managing the rear-view mirror.
Untimely money flow detection
Stock expiry
Weak outlets going unnoticed
Too many tools to manage at once
Margins Leak
During the initial discovery with the client, to understand, here are few set of questions that I asked the stakeholders
I talked to a few customers from senior operations, IT and sales team to understand their pain points and problems , and to understand more about their workflows. Here were some set of questions that I asked around different set of customers.
After conducting interviews I realised that this product should be one platform,with three vantage points. A restaurant group isn’t one user ; it’s a hierarchy, each level needing a different view of the same business. I designed AI-ROS around three primary roles, with the navigation adapting to each.
All locations
Director/admin
lens of this case study
Multiple locations
Region Head
One restaurant
Outlet Manager
The principle held across all of them: an Outlet Manager never sees another region’s numbers, a Region Head never touches company-wide settings, so the nav gives each role only its slice.
User personas
User requirements post conducting user interviews and from user personas.
Post the audit of some of the tools that were exisitng in the market, I realised that every operator I spoke to was already drowning in reports POS exports, labor sheets, reservation dashboards. More data wasn’t the answer; they had too much of it. What they lacked was synthesis , one place that turned all of it into a decision. That’s why AI-ROS leads with answers and alerts, not another dashboard to read.
They already run best-in-class systems, and they’re not giving them up. The gap wasn’t the tools; it was that none of them talk to each other. So AI-ROS doesn’t compete with the stack ,it sits above it, pulling every source into one picture.
POS
Toast · NCR Aloha
Labor
Harri · Deel
Reservations
SevenRooms · Resy · OpenTable
The problem statement
Customers run on disconnected tools, so they find out about expiring stock, margin leaks and weak outlets after the loss. My role is to craft a solutions which bridges all three gaps that can help in making one view with multi role and multi location access easier with help of AI.
86% efficiency in inventory management
Scaled from 0 to 20+ locations
Onboarding that fills the empty room.
A conversational interview , not a settings panel. Each answer pre-wires the product, so the dashboard is never empty.
Dashboard
The Dashboard — the business as one thing. Alerts sit above the metrics, an AI narrative diagnoses what's moving and why, and the P&L, labor plan, and invoice queue all resolve in the same view.
Data Ingestion
The thesis, made literal. Toast, Aloha, Harri and SevenRooms connect directly with CSV as a fallback because the job was never to replace the operator's stack, only to make it finally speak.
Sales Forecasting
The AI makes the call, the operator keeps the veto. Configurable inputs feed a cover and sales forecast, warning bands flag overtime and SPLH risk, and a Manual Override column lets any number be overruled before the forecast is posted.
Labor Scheduling
The thesis, made literal. Toast, Aloha, Harri and SevenRooms connect directly with CSV as a fallback because the job was never to replace the operator's stack, only to make it finally speak.
Approvals
Most tools design only the happy path. I designed all three states of the invoice workflow — the routine it handles, the scan it can't read, and the exception it doesn't yet know — because a system you can trust is one that's honest about what it can't do.
Approvals - the routine
The happy path, handled. Auto-coded invoices clear in a single click, each tagged with its cost category and location the volume work moves without the operator touching it.
OCR Failure - the honest "I can't read this"
When the scan is below the confidence threshold, the system flags it rather than guessing showing the raw extracted text and offering manual coding or a re-process. It would rather admit uncertainty than post a wrong number.
Exception - the failure that teaches
An invoice from a vendor with no matching rule stops the line, explains exactly why, and offers to create the rule on the spot so a one-off human fix becomes permanent automation.
Rules Engine
Automation, dragged into daylight. The logic that codes every invoice is a plain table — vendor equals Sysco Foods → Food — with priorities, on/off toggles, and edit rights sitting with the operator, not the black box.
What could I have done better
With an extended timeline, I could have designed for both the themes light and dark and I could have visually improvised it without the constraint of the current design system.