RetroPilot

Inspiration

A couple of days before the hackathon, we spoke with the president of NOCO, and what he told us stuck with us. Homeowners and building owners in Western New York have struggled for a very long time with the same problem: they want to lower their energy bills and switch to cleaner energy, but they don't know which upgrade to choose, what it will really cost, or who to trust to install it. NOCO has been trying to solve this for years, and they were glad to share a dataset with us to help us build something real.

We want to live in a world where renewable energy is loved and encouraged, and where nobody gives up on it because the process feels confusing. We built RetroPilot to make that pain point disappear.

What it does

You give RetroPilot a few simple inputs about your building, such as its address, type, size, age, fuel type, and energy usage, and you can add photos of the property. RetroPilot then tells you:

Which upgrade to choose. It ranks the best options (heat pumps, insulation, lighting, water heating, solar, and more) and highlights the AI's top recommendation.
How much it will cost and how much you'll save. You see a realistic cost range for Buffalo and Western New York, the estimated annual savings, and the payback period for every option.
Who to call. For every recommendation, RetroPilot suggests installers in the Buffalo area and explains why, aiming for the most affordable option for the job.

How we built it

Frontend: Next.js, React, TypeScript, Tailwind, shadcn/Radix, and Recharts for a simple guided flow.
Photo inventory: users upload photos or capture them on a phone, and a GPT vision model identifies and counts visible appliances, lights, and equipment, with confidence scores. Users review the counts before anything is used.
Custom trained AI energy model: a TensorFlow/Keras model takes the reviewed photo inventory, building details, and utility data to predict electricity and gas baselines.
Upgrade engine: a deterministic TypeScript engine turns those baselines into savings and costs, using NREL PVWatts for solar and an incentive catalog for net costs.
Installer recommendation: the GPT API ranks installers for each upgrade, choosing only from a vetted Western New York installer directory and cost data, so it can't make up companies or prices.
Data: the dataset NOCO shared, plus public sources for costs and incentives.

Challenges we ran into

Photos can't tell you everything. A picture can show how many appliances there are, but not their wattage, age, or how often they run. We handle this by letting users review the counts and by showing savings as ranges, not falsely exact numbers.
Keeping GPT honest. Left alone, a language model will happily invent installer names and prices. We constrained it to choose only from a verified directory and validated every ID it returns.
Not double-counting savings. Upgrades like insulation and heat pumps overlap, so combining them needed careful logic.
Building a lot in one hackathon. We went from first conversation to a working product in just a one days.

What we learned

Real problems become much clearer when you talk to the people living with them, like NOCO.
Pairing AI with deterministic, auditable calculations gives better results than relying on AI alone.
Being transparent about uncertainty builds more trust than claiming perfect accuracy.

What's next

Expand the installer directory with verified quotes and reviews across Western New York.
Validate and retrain the prediction model on more real-property data.
Connect to live incentive and utility program data.
Partner with local companies like NOCO to turn recommendations into real installations.