Demand, seen before it happens.
A machine-learning forecasting platform that ingests historical sales, seasonality and live signals to predict demand — so teams plan inventory, staffing and spend with confidence instead of guesswork.
Fresh food forgives no one’s guesswork.
A fresh-meal operation lives or dies by tomorrow’s number: cook too much and margin goes in the bin, cook too little and fridges sit empty at lunch. Planning ran on experience and spreadsheets — good instincts, no memory, and no way to scale across a growing fridge network.
We built the platform that remembers everything: two years of sales history, seasonality, weather and live signals from the fridge fleet, distilled into per-site, per-dish demand forecasts. Inventory, staffing and spend now plan against a number that has earned its trust — 98% accuracy in production.
In production the model plans the day before the kitchen starts it: 98% forecast accuracy, 58% fewer stockouts and 41% less excess stock — with drift monitoring and scheduled retraining keeping it honest as demand shifts.
Instinct doesn’t scale.
Good planners ran the operation on experience — which can’t be copied to the next ten sites.
Tomorrow’s number was a guess.
Fresh food punishes both directions of error, every single day.
- Overproduction — margin literally binned at the end of each day.
- Stockouts — empty fridges at lunch are revenue and trust lost.
- Spreadsheet memory — planning had no systematic recall of seasonality.
- Instinct is local — expertise stuck with one planner at one site.
A number that earned its trust.
Per-site, per-dish demand forecasts distilled from history, seasonality and live fleet signals.
- 98% forecast accuracy — production-grade, measured continuously.
- Live signals in — sales history, weather, seasonality and fridge telemetry.
- Planning views out — inventory, staffing and spend against one number.
- Self-correcting — automated retraining keeps the model honest as the network grows.
From signals to a plan.
The full loop — ingestion, prediction, planning views and the retraining that keeps it honest.
Signal ingestion
Sales history, seasonality, calendar effects and live fleet telemetry flowing through engineered pipelines.
Demand forecasting models
Per-site, per-dish predictions tuned to the horizon planners actually buy against.
Inventory planning views
Forecasts translated into order quantities and production plans — not raw charts.
Staffing & spend planning
The same demand curve drives shift planning and purchasing budgets.
Drift monitoring
Accuracy tracked continuously against reality, with alerts when the world shifts.
Scheduled retraining
Models retrain on fresh data automatically — accuracy survives next quarter, and the one after.
Earning the planners’ trust.
Four phases — including months of running silently next to the humans.
Conceptualisation
Framed the cost function with operators: a binned meal versus an empty fridge.
Design
Planning views that say what to cook — not dashboards that say “it depends”.
Development
Ingestion, the feature pipeline, the models and the automated retraining loop.
Deployment
Shadow-mode trial against human plans, then site-by-site cutover.
What kept us up at night.
The problems that decided whether the product worked at all.
Trust before automation
Planners ignore a black box. The model ran in shadow mode against human plans until its track record — not its math — won the argument.
Cold-start sites
New locations have no history. Transfer from similar sites gives day-one forecasts that converge as local data arrives.
Drift in a changing network
Menus rotate, sites open, seasons shift. Automated retraining and accuracy monitoring keep 98% honest over time.
Tech stack.
A learning loop over the whole fridge network.
Numbers the owners watch.
The platform paid for itself in the first quarter — and then got extended to demand planning.
A production number, measured against what actually sold — not a holdout-set brag.
Empty-fridge moments collapsed; lunch is there when the customer is.
Less fresh food cooked into the bin — margin and sustainability, same lever.
“The forecasting platform paid for itself in the first quarter. Six months later we asked CODT to take the same approach to demand planning — they treat both like one product.”
Keep exploring.
Every project here is live, paid for, and earning revenue for its owners.
Have a problem worth
solving well?
Tell us about your product, your timeline and your constraints. We reply within one business day with an honest read on fit, scope and the right team.