The math behind every agent decision — solved in minutes
When an agent needs a plan, it hands the problem to mathematics: millions of combinations, every capacity, material, due-date and routing constraint, and the lowest-cost or best-service answer.
Layer 2 · Optimization Engine
Real math, not guesswork: the best feasible plan in minutes
Language models are great at language. They are not solvers. When an agent needs a plan, it calls an optimization engine that can prove its answer satisfies every constraint.
Convect’s in-house engine turns planning into a large-scale mathematical optimization model:
- Decisions: what to make, where, how much, when to start, with which materials, shipped to which DC…
- Objective: minimize total cost (production + inventory + transport + shortage penalties) or maximize service
- Constraints: capacity, materials, due dates, routings, batch sizes — all drawn from the ontology
The engine searches millions of possible plans in minutes. Where problems are small enough it proves global optimality; at very large scale it returns a high-quality feasible plan with a measured gap to the theoretical bound.
Optimization model (simplified)
min Σ production + inventory + transport + shortage penalty
s.t.
Σp xp,l,t · hoursp,l ≤ capacityl,tcapacity
Σp BOMm,p · xp,·,t ≤ availablem,tmaterials
finisho ≤ dueo + lateodue dates
changeoverp→q,l ≥ cleanp,qprocess rule (from ontology)
Objective convergence
IllustrativeCost structure: before vs. after
Illustrative · before = 100Scale
Multi-plant × multi-line × multi-SKU × multi-period models solved together
Speed
Tens of millions of plan options evaluated and compared in minutes
Incremental replans
Only the affected part of the plan is re-solved when data changes
Concurrent planning
Demand, production, materials, inventory and transport optimized together — not in sequence
Multiple objectives
Lowest total cost, highest service, or a weighted mix — you decide what matters
Explainable
Binding constraints and bottlenecks surfaced, so planners see why the plan looks the way it does
What this means for you
Your planners used to spend a whole day iterating in spreadsheets. Now the engine returns the best plan in minutes — with every constraint respected. The planner’s job shifts from doing arithmetic to making decisions.
Comparison
Optimization vs. rule-based scheduling
Rules give you a plan that works. Optimization tells you what the best plan is.
| Rules / heuristic scheduling | Convect optimization engine | |
|---|---|---|
| Goal | A feasible plan, built from rules | The best plan under all constraints |
| Scope | Production, materials, transport planned in sequence | Production, materials, inventory and transport optimized together |
| Trade-offs | Tuned by reordering rule priorities | Cost and service weighted explicitly |
| When things change | Full re-run, manual comparison | Incremental re-solve of the affected scope |
| Explainability | Rules are readable, but is it any good? | Binding constraints and bottlenecks identified, improvement quantified |
Applications
One engine across the whole planning chain
Master production scheduling
Capacity allocation and balancing across plants and periods
Finite-capacity scheduling
Line- and workstation-level sequencing, down to the hour
Material requirements planning
BOM explosion and purchasing optimized together
Network design
Site selection, service assignment and flows solved globally
Replenishment & deployment
Replenishment and transfer decisions across multi-echelon networks
Transport planning
Optimal moves under truck types, minimum loads and rate breaks
At Yili, the embedded network optimization engine evaluates tens of millions of plan options in minutes — cutting master planning from 5 days to half a day and adding 5% total capacity.Read the story →
Frequently asked questions
Do you use open-source or commercial solvers?+
Convect built its own optimization engine for supply chain planning (multi-period network models plus specialized algorithms), and combines general-purpose solving techniques depending on the problem. Customers don’t need to buy a separate solver license.
Can you still guarantee optimality on very large problems?+
Where problems are small enough to prove optimality, the engine returns the global optimum. At very large scale it returns a high-quality feasible plan with a measured gap to the theoretical lower bound — and every constraint is always satisfied.
Will planners understand the results?+
Yes. The system highlights binding constraints and bottleneck resources and supports scenario comparison and what-ifs, so planners can see why the plan looks the way it does before making the final call.
Run the optimizer on your own data
Send us a spreadsheet of a recent schedule and we’ll show you how much room for improvement the engine finds.