Machine learning, mixed-integer linear programming, optimisation, and the occasional bit of signal processing nobody else wanted to attempt. Applied to real energy problems, in production, where the maths usually has about 200 milliseconds to be right.
Applied maths in energy isn't a paper you present and forget. The output goes off and does something — it dispatches a battery, schedules a plant, or turns up as a number on someone's electricity bill that they will absolutely query. That changes how you build it: accurate enough that you'd stake money on it, fast enough to run on whatever modest hardware is actually in the field, and never so clever that nobody can tell you why it did what it did.
“Electricity demand isn't a number to be met — it's a signal with layers. Read it the right way and the optimisation almost designs itself.”
Turn the business problem into the right mathematical one — objective, constraints, data.
MILP, LP, ML, signal processing — fit the technique to the problem, not the fashion.
Develop the model and test it against real data and the optimal benchmark.
Engineer it to run at the speed the application needs — milliseconds, on modest hardware.
Put it into production with the software and data platform around it.
Maths that runs in production, not just in a paper.
Decomposing demand by timescale to dispatch 72 generators in 0.2 ms on a Raspberry Pi — within 2.3% of optimal, beating merit-order by 36–42% on cost.
View project →A patent-approved method (US 10,734,809) that infers fixture run-times from a single circuit signal — accurate enough to put on a bill.
View project →Mixed-integer linear programming turns biological, storage and market constraints into a 15-minute dispatch schedule against day-ahead prices.
View project →Demographic-driven demand personas and Time-of-Use modelling that size infrastructure and flatten peaks.
View project →Sometimes you should — but the hard problems don't fit. A 72-generator fleet has more on/off states than grains of sand on Earth; we design methods (like wavelet-based dispatch) that sidestep the blow-up and run in milliseconds on cheap hardware.
When it has to be, yes. Our patented load-disaggregation method calculates the probability of each load being on accurately enough that customers and utilities will sign off the savings on every M&V project.
Yes. We've delivered dispatch as a 90KB dependency-free binary running on a Raspberry Pi — because in the field, fast and lightweight beats elegant-but-heavy.
Both. We prototype and benchmark novel methods against the optimal solution, and we engineer the winners into production systems with the software and data platform around them.
We formulate it, solve it, and make it fast enough to run for real.