Never let the LLM do the arithmetic
In GridWise, a language model reads the operator's notes, but an exact linear program does every calculation. Here's why that split works.
GridWise plans a campus’s electricity for the next 24 hours: when to buy from the grid, when to use solar, and when to charge or drain the battery. An operator can steer it with notes in plain English:
The data center requires at least 80 kWh to remain in the battery from 6 PM until 10 PM. Grid intake must stay at or below 190 kWh from 7 PM until 10 PM while the substation is constrained. A seminar room booking was moved to next week.
The tempting design is to hand everything to an LLM. GridWise doesn’t, and that’s the point of this writeup.
Language for intent, maths for numbers
An LLM is good at reading the notes and bad at guaranteeing that a schedule obeys physics. So the work is split:
- The LLM classifies each note into one of six typed directives, and nothing else. The notes above become:
{"type": "minimum_battery_reserve", "applies": true, "hours": [18, 19, 20, 21], "minimum_energy_kwh": 80.0}
{"type": "max_grid_window", "applies": true, "hours": [19, 20, 21], "max_grid_kwh": 190.0}
{"type": "no_op", "applies": false}- Guardrails check the output before it can influence anything: hours are unique and in range, factors sit between 0 and 1, and percentages become kWh.
- A linear program does the numbers. 96 variables (grid, solar, charge and discharge for each hour) and 73 constraints go to the HiGHS solver, which returns the exact cheapest plan in under 5 ms.
The seminar note is a distractor. It becomes a no_op, and the plan ignores it.
Why the split holds up
- Every plan is physically valid: energy balance, battery limits, reserve floors and an end-of-day battery level equal to the start. These are constraints, not suggestions the model might forget.
- The LLM can fail safely. If it’s slow or rate-limited, a deterministic extractor takes over, and the guardrails treat both the same way.
- It’s testable. 97 tests replay the public sample cases and adversarial notes, and check every returned plan against the physics.
The general lesson: let the model translate, and let something exact compute.