Skip to Main Content

Solar Proposal Agent

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Demo video · www.loom.com/…

Solar Proposal Agent turns a site's documents (datasheets, surveys, notes) into a solar + storage design in which every number cites its source, and hands whatever it cannot prove to a person. The problem: in 2024 I wrote a proposal to keep a rural water plant and an evacuation footbridge running through power cuts in Valle Hermoso, Ecuador. It was set aside before anyone reviewed it. Rereading it, I saw why it was weak: the system was never sized and the budget totals did not add up. Solar plus storage is the obvious answer for communities like this; what blocks it is the technical proposal. This tool is the one that proposal needed. How it works: Muse Spark 1.3 runs an agent loop with tools. It reads the case documents (text and PDF datasheets) and records inputs. A value taken from a document must quote that document exactly and appear inside the quote; code checks this, and an invented quote is rejected. Deterministic code, not the model, does all calculations (pumping energy, lighting, PV array, battery bank, CO2 avoided from IPCC-derived factors) and fetches the site's worst-month irradiance from NASA POWER. The proposal text is written by code, so it cannot contain a figure that was not computed. Anything unproven becomes an open question for a person. Assumptions are flagged, and when site facts such as water demand or pump head are only placeholders, the report is stamped ILLUSTRATIVE ONLY. What I tested live: a synthetic example case, sized end to end with a result that matches a hand calculation, and a real product datasheet with notes from my 2024 proposal, where the agent declined to invent a design and returned eight specific questions about what must be measured. Live runs exposed defects (percentages recorded as 45 instead of 0.45, a status-tool loop, a run reported as finished without sizing); I fixed them and added tests (44 offline tests). Honest limits: the worked example is synthetic. The agent does not yet recognise luminaires that carry their own panel and battery, compare products, or price a budget. Photo reading is implemented but is not part of the demo. It produces a preliminary design for a qualified engineer to review and sign.

Team