Solar Proposal Agent: AI That Never Invents a Number

In 2024 Ecuador rationed power for up to 14 hours a day. Solar plus storage is the obvious answer, but a proposal has to be sized, priced and defended, and communities applying for funds rarely have an engineer. I wrote one myself in 2024 for a rural water plant and its evacuation route. It was never funded. Reading it again, the weak spot was the technical basis: the system was never sized and the budget did not add up. This agent is the tool that proposal needed.

It reads a site's documents (datasheets, surveys, photos) and produces a preliminary solar + storage design where every number traces back to its source. Anything it cannot prove is handed to a person.

Muse Spark 1.3 (Meta Model API) runs the agent loop: it reads the documents, records inputs and decides what to ask a human.

Quote-checked inputs: a value taken from a document must quote it exactly, and the value must appear inside the quote. Invented quotes are rejected in code.

Deterministic sizing: pumping, lighting, PV array and battery are calculated by code, not by the model. Solar resource comes from NASA POWER data.

The proposal text is written by code, so it cannot contain a figure that was never computed.

Open questions: assumptions and anything measured only from a photo are listed for a person to confirm before signing.

The worked example is a synthetic case (it sizes about 7.15 kWp of PV, 13 modules) that exercises the pipeline. It is not a real site.

On a real product datasheet, the agent asked what it could not prove instead of inventing a design. Testing also exposed limits I documented openly, such as autonomous luminaires not being recognised yet.

It is a preliminary design for a qualified engineer to review and sign. It does not price from real quotes or design the electrical layout.

41 offline tests. Code is open source (MIT).

Designed and built the full agent end to end: tool-calling loop, source-checking rules, sizing code, report generation and tests.