Solar Copilot: one app from buying rooftop solar to running it well

Solar Copilot takes a household from choosing rooftop solar to getting the most out of it, built for hot, dusty places with load-shedding and slab tariffs like Pakistan. One home profile feeds six tools: system plan, tomorrow's appliance plan, cleaning alert, quote checker, bill slab coach and backup compare. AI reads and explains; tested code calculates. Demo video: https://youtu.be/E7hz4AyvZcY

Live demoGitHub1

Overview

Rooftop solar is booming in Pakistan, but households decide blind: which panels, how to place them, when to run appliances during load-shedding, and whether a quote or a bill is fair. Solar Copilot answers all of that from one home profile, kept only in the browser (no account, no database).

Six tools: (1) Plan my system: tilt and direction from EU PVGIS, how many panels fit the roof, row spacing so rows never shade each other in winter, mounting height, size for your bill, and panel technology for your heat and dust. (2) Tomorrow's plan: an hour-by-hour appliance plan from the Open-Meteo solar forecast, battery and outage schedule. (3) Cleaning alert: dust loss vs wash cost, and whether coming rain will clean the panels. (4) Quote checker: flags undersized or mismatched parts, missing protection and prices outside normal ranges, with the rule behind each warning. (5) Bill slab coach: explains your tariff slab and how many units to save to drop one. (6) Backup compare: generator vs UPS vs solar + battery, cost per outage hour and payback.

Why you can trust the numbers: the AI only reads text and explains; every number comes from tested Python (233 tests). Each result shows its inputs, assumptions and sources; tariffs, prices and rules live in versioned files and are editable in the app. Quotes and bills are treated as untrusted input.

Data: PVGIS (EU JRC), Open-Meteo forecast / air quality / archive, Ember grid CO2 via Our World in Data, NEPRA tariff S.R.O. 279(I)/2026.

Demo video: https://youtu.be/E7hz4AyvZcY

How it uses Vultr

Vultr Cloud Compute VM: runs Coolify, which builds the app from the public GitHub repo and serves the Docker Compose stack (nginx serving the React frontend with security headers, proxying to a stateless FastAPI backend). No database: the home profile stays in the user's browser and is sent with each request. Vultr Serverless Inference (model mimo-v2.6-pro, OpenAI-compatible API, JSON mode): the only AI in the app. It turns messy text (appliance lists, installer quotes, electricity bills) into schema-validated data, and writes short plain-language summaries of results the Python code already calculated. Timeouts, 2 retries on invalid JSON, and a guard that hides any AI text containing a number the code did not compute. We tested every chat model on our key for reliable JSON before choosing this one. Why: one small VM keeps hosting cheap and simple for a stateless app, Coolify gives push-to-deploy, and Serverless Inference gives an LLM without running a GPU.

1 comment

Sorted by votes

Wow Amazing guys