Inspiration
I noticed a weird disconnect: we track our steps, our calories, and our bank accounts, but we have no idea about our environmental debt. Most people don't know that their morning coffee or the shirt on their back cost thousands of liters of water to produce. I wanted to build something that makes that "invisible" cost impossible to ignore. As a solo dev, I wanted to see if I could use AI to turn a simple camera into a powerful tool for climate transparency.
What it does
BluePrint is a "Point-and-Shoot" for sustainability. You take a photo of any product, and within three seconds, the AI breaks down its environmental footprint. It shows you the water usage (broken down by type), the carbon emissions, and—crucially—a better alternative. It’s not just about showing you what’s wrong; it’s about showing you how to fix it through "Smart Swaps" and personal challenges. All verified by Water Footprint Network global.
How I built it
Being a team of one, I had to be smart about my tech stack. I used Gemini 2.5 Flash as the "brain" because it can handle vision and data synthesis in one go.
- Backend: I used Python with Pydantic to make sure the AI’s data was actually accurate and structured before it hit the UI.
- UI/UX: I pushed Streamlit to its limits with custom CSS to get that frosted look you see in modern apps simple and effective for attention.
- Data: I grounded the AI's logic in real-world stats from the WFN 2024 and IPCC databases to make sure the "truth" it gives is backed by science. Ref
Challenges I ran into
The hardest part was working in five places at once. When the JSON parser broke at 3 AM, there was no "backend guy" to fix it—it was just me.
- The Parser Struggle: Getting an AI to consistently output perfect code for charts is like herding cats. I had to write a custom, fail-safe extraction system to handle the vision-to-data pipeline—twice.
- Accuracy: I didn't want this to be a "random number generator." Tuning the prompts with enough reference data to get 85%+ accuracy across different product categories was a massive trial-and-error process, used AI-assisted tools for this generation ai was not mentioned to be forbidden so i utilised it.
Accomplishments that I'm proud of
I’m really proud that I managed to get the analysis time under 3 seconds(no dead loading). Usually, vision models are slow, but I optimized the pipeline so it feels instant. I’m also proud of the.......UI—I didn't want it to look like a boring spreadsheet; I wanted it to feel like a app people actually want to use.
What I learned
I learned that Multimodal AI is a game-changer for solo developers. It allowed me to build features that would have taken a team of five just a few years ago. I also learned that once you see the "water price" of a steak or a new phone, you can never really look at shopping the same way again.
What's next for BluePrint
I want to add Barcode Scanning to get 100% precise data from specific brands. I’m also looking at building a browser extension so that while you’re shopping on Amazon, BluePrint can show you the environmental cost before you even hit "Add to Cart."
Log in or sign up for Devpost to join the conversation.