CarbonIQ — Carbon Footprint Tracker
AI-powered carbon footprint tracker where you just type what you did — "drove 45km, ate 1kg of beef" — and Gemini figures out the rest. No dropdowns, no friction.
Case Study
The Problem
Carbon trackers die at the input step. Nobody wants to fill out a form with five dropdowns every time they eat lunch or take a bus. The friction kills the habit before it starts.
My Approach
Built a single natural-language input powered by one Gemini 2.0 Flash call that handles extraction, calculation, and advice simultaneously. Chose single-prompt over 3-agent pipeline to reduce API calls on free-tier limits and eliminate silent failure points. Added a confidence-gated confirmation step to protect data integrity.
The Result
A fully functional carbon tracker with real AI, real auth, and real Firestore data — not a demo. Includes Earth Debt clock, Carbon Passport with 10-year projections, personalized daily challenges, and streak/badge system. Built for PromptWars Virtual — Main Challenge 3.
What It Does
- Natural language + photo input — AI extracts activity, quantity, and emission factor automatically
- Real-time "Earth Debt" clock on homepage ticking up in live rupees
- Carbon Passport page with 10-year projection chart comparing two futures
- Personalized daily challenges based on your actual worst-emission category
- Streak tracking, badges, and CO2 history via Firebase Firestore
- Confirmation step when AI confidence is low — never saves bad data silently
Challenges & Solutions
Challenge 1
Most carbon trackers fail at the input step — users abandon complex forms.
Solution
Replaced all dropdowns with a single natural language field. One Gemini 2.0 Flash call extracts activity, quantity, unit, and emission factor in a single pass — no multi-agent overhead.
Challenge 2
Preventing garbage data from silently entering user carbon history when AI misreads input.
Solution
Added a confidence-based confirmation step — if Gemini is uncertain, it surfaces what it detected and waits for user approval before touching Firestore.
Challenge 3
Making abstract CO2 numbers feel real and motivating to non-expert users.
Solution
Built relatable comparisons ("= 50 phone charges"), a live Earth Debt rupee clock, and a terminal-styled Carbon Passport page that deliberately feels different from the rest of the app.
Project Gallery
What I Learned
Single-prompt AI architecture vs. multi-agent tradeoffs, Firebase Auth debugging (CSP header conflicts), Recharts data visualization, India-calibrated emission factor modeling