Six engines compute the answer — NEC load calculations, conductor and breaker sizing, charge curves, depot load management, site economics, tariff selection. Twelve Claude instruments then explain it, audit it, and defend it to an inspector. The arithmetic is checkable. The reasoning is on the record.
Whether a panel can carry a charger is a calculation, not an opinion. So is the conductor size, the charge time, the depot peak, and whether a site ever breaks even. These get guessed at every day, and the guesses cost thousands.
Every engine is pure arithmetic against published code tables. Same inputs, same answer, every time — no model in the path. Each one ships with unit tests you can run from the repo.
Then the AI layer: every engine result can be sent to Claude for a written review — what the number means, what an inspector will question, what to change, and the three things the model would check next. The computation is deterministic; the interpretation is where the model earns its place.
Panel rating, floor area, vehicle, dwell window, tariff. A project holds it all and every engine reads from the same set of facts, so nothing is re-typed and nothing drifts out of sync.
Line-by-line arithmetic with the governing code section printed beside each step. Charts render the charge curve, the load profile, the break-even. Nothing is hidden behind a model.
The computed result goes to an instrument with a domain-expert brief: what it means, where it is fragile, what an inspector will ask, and what to do next. Export the whole thing as a project report.
Not one chat box with twelve prompts. Each instrument has a fixed role and a specified output shape — a comparison declares its weights before it scores; a scenario model tables every assumption and names the one that breaks it. Output from any instrument can be handed to any other.
Arithmetic against a public code table should not be behind a paywall. What you pay for is the interpretation layer, saved work, and everything a team needs.
The whole deployed surface is a static front end, two edge functions and a managed Postgres. There is very little to attack, and we can describe all of it in one screen.
npm test in
the repo and 55 assertions check the output against hand-worked examples.
Claude sits on top to interpret, critique and write up what the engine produced.
That order matters: a model that guesses a conductor size is a liability, and a
model that explains a computed one is genuinely useful.Six engines, twelve instruments, no sign-up to start.