HostGraph: a margin-leak detection command center for restaurant groups

HostGraph is a procurement intelligence dashboard built for multi-location restaurant groups — the kind of operation where margin is decided not at the register but in the back office, one purchase order and one delivery invoice at a time. It pulls vendor pricing, order history, and shrinkage into a single command center so a controller or ops lead can see a margin leak while it's still small.

The problem

Restaurant groups don't usually lose margin in one dramatic event. They lose it quietly. A produce vendor's price creeps up 4% over a quarter and nobody reconciles the invoice against the last order closely enough to catch it. Shrinkage — spoilage, waste, theft — happens at the store level and rarely gets tied back to purchase volume, so it never shows up as a number anyone is accountable for. Reorder decisions get made from memory and gut feel by whoever is on shift, not from a baseline of what the location actually needs.

None of this trips an alarm. It shows up months later as a margin line that's drifted the wrong direction, with no clear path back to a cause. By the time someone notices, the leak has already compounded across every location in the group.

The approach

HostGraph is built around the idea that margin leaks are traceable if you're watching the right signals continuously, instead of reconstructing them after the fact from a P&L. That means four core capabilities, all pointed at the same underlying data:

• Vendor price-trend analysis — tracking what each vendor is actually charging over time, so a 4% creep shows up as a trend line, not a surprise at quarter-end
• Reorder logic — flagging when a location's purchase pattern drifts from its own historical baseline, catching over-ordering and under-ordering before either one hits the P&L
• Shrinkage tracking — tying spoilage, waste, and unexplained loss back to specific SKUs and vendors instead of leaving it as an unattributed rounding error
• Alerting — surfacing anomalies the moment they cross a threshold, rather than waiting for someone to go looking for them

The goal isn't a bigger dashboard. It's a smaller number of things that actually get looked at, because the system is the one doing the watching.

Architecture

HostGraph ships as six dashboard views, each a different lens onto the same margin engine rather than six disconnected tools bolted together. Illustratively: a vendor overview for at-a-glance spend and standing, a price-trends view for the creep analysis, a reorder-recommendations view built on the drift logic, a shrinkage log tying losses to SKUs and vendors, an alerts feed for anomalies as they fire, and a rollup summary view for the cross-location picture. Every view reads from and writes back to the same margin engine, so a price change flagged in one view is immediately reflected in the reorder math and the alert thresholds in the others — there's no separate data model per screen to keep in sync.

The stack is React 19 on Vite with Tailwind for styling, and the whole build is a clean TypeScript compile verified end to end — every view, every data path exercised, not just the happy path spot-checked in a browser. That matters because HostGraph is meant to be handed to an operator and actually run, not just demoed once and left as a prototype.

Results

HostGraph is demo-ready: a complete, working system spanning the full margin engine and all six views, built to be walked through end to end rather than shown as static mockups.

HostGraph — six views, one margin engine

Lessons learned

The most useful realization building HostGraph was that the analytics were never the hard part — vendor price trends, reorder drift, and shrinkage-by-SKU are all fairly standard aggregation problems once the data is clean. The hard part, and the part that actually determines whether margin gets saved, is the alerting layer. A dashboard nobody checks doesn't stop a leak, no matter how accurate the underlying numbers are. It was the alerting layer that turned "the data exists somewhere in the system" into "someone actually saw it in time to act." Building HostGraph as alert-first rather than dashboard-first, and treating the six views as places to investigate an alert rather than places to go looking for problems cold, was the design decision that made the whole thing worth using.