Optimisation
Scheduling, allocation, routing, design and packing. We use mixed integer programming where it scales, and custom heuristics where it doesn't.
Mathematical optimisation that works on real industrial problems.
Custom algorithms, including for problems off-the-shelf solvers can't finish.
We turn your scheduling, allocation, routing and planning decisions into mathematical models, then build the software that solves them. Our methods are published and tested on real operating data.
Problems we solve
If a decision has an objective and constraints, we can probably model it.
Scheduling
Production lines, shifts, maintenance and project plans.
Allocation
Staff, machines, budgets and limited materials.
Routing and flow
Fleets, deliveries and supply networks.
Location and network design
Warehouses, plants, service points and infrastructure.
Cutting and packing
Material cutting, and container and pallet loading.
Planning under uncertainty
Inventory, capacity, and policies tested by simulation.
Published results
What our methods have shown so far, much of it in logistics. Figures are as reported in the linked papers.
-
≈1%
arXiv 2610.30139
Going further than a time-limited exact solver
On logistics network design, our variable neighbourhood search stays within about 1% of the best dual bound on small networks. On medium networks it is up to 55% better than a time-limited exact solver.
-
5.0%
arXiv 2606.09455
Finding savings by simulation
Replaying a month of real parcel flow in a discrete-event simulator showed one production policy could save 5.0% of cost at the same lead time. The fix changes two parameters and needs no capital.
-
20%
Procedia CS, 2014
Handling real-life constraints at scale
Truck and trailer routing on real retail data, with about 400 stores and 100 vehicles per warehouse, plus time windows and split deliveries. Our heuristic found solutions 20% cheaper than a simple greedy method.
-
4×4
arXiv 1712.08437
Modelling the structure of search itself
A mathematical programming model of the best network for decentralised nearest neighbour search. We solved it exactly on a 4×4 lattice and found heuristic solutions up to 7×7.
-
NP
arXiv 2007.05020
Establishing how hard a problem is
We modelled team selection in the game Dota Underlords as an integer program, solved it on real data, and proved the problem NP-complete.
What we build
Every project is custom, from the first formulation to working software.
Simulation and agent models
Replay your real operating data, sweep policy settings, and see the trade-offs before anything changes in production.
Vector search
Better approximate nearest neighbour search, from graph structure to tuning on your data for recall, latency and memory.
Machine learning and vision
Forecasting, detection and inspection models, including the demand and failure predictions that optimisation depends on.
Real costs come in steps
Machines, shifts, batches and vehicles come in whole units, so a small extra demand can mean paying for one more. A linear model can't see that jump. Move the slider to compare them.
- Units needed
- –
- LP relaxation
- –
- Cost the LP misses
- –
Example with a capacity of C = 20 per unit.
The general model
Pay for each unit, and for the work it does
Every demand is met
Enough whole units are running to carry the work
How a project runs
We start with your operation, and you keep everything we build.
Formulate
Write down the objective, the constraints and the exceptions with the people who run the operation.
Measure a baseline
Score your current process on historical data, so every result has a fair comparison.
Solve
Build the algorithm, then test it against the baseline and a mathematical bound.
Deploy
Connect it to your systems and hand over the code, the model and the documentation.
Papers
Open to read, including the formulations and experiments.
- 2026A General Variable Neighborhood Search for Logistic Network Flow Design with Nonlinear Edge Costs and Transshipment CapacitiesKovaleva, Ponomarenko
- 2026A Benchmark Generator of Realistic Logistic Transport Networks with Nonlinear Edge Costs and Transshipment CapacitiesKovaleva, Ponomarenko
- 2026Tuning Dispatch Thresholds for Fixed Last-Mile Routes: A Simulation-Based Pareto Analysis of a Production PolicyPonomarenko, Antonov
- 2021Overlapping community detection in networks based on link partitioning and partitioning around medoidsPonomarenko, Pitsoulis, Shamshetdinov
- 2020Dota Underlords Game is NP-CompletePonomarenko, Sirotkin
- 2017A Model of Optimal Network Structure for Decentralized Nearest Neighbor SearchPonomarenko, Utkina, Batsyn
- 2014Heuristic for a Real-Life Truck and Trailer Routing ProblemBatsyn, Ponomarenko