FuelCast
A demand-forecasting and business-intelligence platform for fuel retail — predicting fuel and shop demand, catching margin leaks, and answering plain-English questions from operators.
Client — Shell South Africa — multiple forecourts
Client anonymized. Every figure and chart on this page uses synthetic data calibrated to look realistic — we never publish a client's live numbers.
Fuel forecourts run on thin margins and perishable stock. FuelCast forecasts fuel and shop demand up to 90 days out, turns it into ordering decisions, and puts a plain-English AI assistant in front of the whole thing so a manager can just ask.
Demand forecasting & shop ordering
Forecast demand
281k L
Gross margin
22.4%
Deadstock flagged
R 38.2k
Data quality
99.2%
[ DEMAND FORECAST ]
Petrol 95 · liters / day
[ STOCK ORDERING ]
Recommended · next 7 days
- Beverages1240 u
- Snacks880 u
- Bakeryspoilage risk560 u
- Energy420 u
- Delispoilage risk340 u
[ ASK FUELCAST ]
Answers from the forecast · figures synthetic
Interactive / built with synthetic data for this page
[ THE CHALLENGE ]
A fuel retailer's real signal is buried: pump sales, shop tills, and card terminals reconcile only at month-end, stock spoils before anyone notices, and pricing is regulated to the cent. The operator needed to see demand coming, order the right stock, and understand why — without hiring a data team.
How we built it
From signal to shipped.
Foundation-model forecasting
A TimesFM 2.5 time-series model forecasts fuel demand per grade and shop demand per category, with South-Africa-specific covariates — paydays, fuel-price change days, holidays, weather. Always benchmarked against an honest seasonal-naive baseline; the model roughly halved the error.
Decisions, not just charts
Forecasts feed a stock-ordering engine — newsvendor logic for perishables, safety-stock for packaged goods — so the output is 'order this much', not a graph a manager has to interpret.
An operator's BI dashboard
A themed dashboard surfaces demand, profit & loss, deadstock, and data-quality — with reconciliation gates that flag bad point-of-sale entries instead of trusting them blindly.
Ask, in plain English
An in-dashboard AI assistant routes questions to read-only tools (SQL, the forecaster, the calc engine), never invents numbers, and stamps every figure. Non-technical staff get answers without learning the tool.
What shipped.
Fuel demand forecast per grade with quantile bands, ~2× tighter than the naive baseline.
Six shop categories forecast and converted into concrete stock-order quantities.
Live ingestion on a 10-minute cycle, deployed and running unattended.
A plain-English assistant that lets any operator query the business directly.
[ NEXT PROJECT ]
Outreach Engine