Planners build the weekly production schedule by hand in Excel. Trade-offs between labour, changeover cost, material on hand, and demand are too many to hold in your head — the schedule that ships is rarely the cheapest one.
The AI approach
Forecast + optimize. ML forecasts demand and yield; a MILP solver finds the schedule that minimises cost subject to capacity, labour, material, and changeover constraints; LLM narrates the plan.
The outcome
~−6.4% cost
Total weekly production cost down ~6.4% on pilot lines. Planner reviews the optimised plan + 2 alternatives instead of building from a blank sheet.
Try itInput → Process → Output
Input — demand + capacity + materials
Demand forecast · this week
SKU CL-117
3,400
SKU CL-204
920
SKU CL-308
1,750
SKU CL-411
2,100
Capacity
Line A: 40h · Line B: 32h (downtime Tue 4h)
Line C: 40h · operators 3 / shift
Click Run demo to forecast demand and solve the optimal schedule.
Optimised schedule MILP · 0.8s
shift
Mon
Tue
Wed
Thu
Fri
Line A
CL-117
CL-117
CL-411
CL-411
CL-411
Line B
CL-204
— ↓
CL-204
CL-308
CL-308
Line C
CL-308
CL-117
CL-117
CHG
CL-411
Labour฿2.14M
Material฿4.86M
Changeover (3 events)฿270k
Total cost฿7.27M · −6.4% vs hand-built
Why this schedule: Tue Line-B downtime forced CL-204 onto Line C, but the solver kept CL-117 on Line A all of Mon-Tue to avoid a costly A↔B changeover. Two alternative plans (within 1.2% cost) attached for planner review.
Three AI types in this use case
SymbolicOptimization solver (mixed-integer linear program): minimise cost subject to machine capacity, labour hours, material on hand, demand met, changeover penalties, inventory caps.
PredictiveDemand forecast per SKU × week; machine downtime predictor; quality yield model per (machine × SKU).
GenerativeLLM writes the plan rationale ("line A on Tuesday despite higher labour because changeover from line B costs ฿180k") and explains tradeoffs to the planner.
The stack
Forecast · Prophet / DeepAR
Solver · Gurobi / CBC / OR-Tools
Constraints · YAML config per plant
Sink · MES / shop-floor schedule
When this works
Constraints are well-codified (capacity, labour, material, changeover)
Demand forecast accuracy is reasonable (MAPE < 25%)
Planner retains override — the solver proposes, planner disposes