# "EcoSphere AI: Why I separated 'logic' from 'AI' when building a carbon footprint assistant"

> Source: <https://dev.to/ajx1tech/ecosphere-ai-why-i-separated-logic-from-ai-when-building-a-carbon-footprint-assistant-1m6a>
> Published: 2026-06-21 05:48:30+00:00

When you say "build a carbon footprint app," the obvious move is: collect some inputs,

ask an LLM "how much CO2 does this produce," and display whatever it says.

That's a trap. LLMs are not reliable calculators. Ask the same model the same emissions

question twice and you'll get two different numbers. For a "track your carbon footprint"

tool, that's disqualifying — the entire premise of tracking requires consistent,

auditable numbers.

EcoSphere AI splits the problem in two:

**1. A deterministic carbon engine** — pure TypeScript functions using documented

emission factors (diet type, transport mode, energy use, shopping habits). Same input,

same output, every time. Fully unit tested. This is what computes your actual footprint,

your Carbon DNA archetype, your Risk Score, and your ranked recommendations.

**2. Google Gemini AI** — used only for what LLMs are actually good at: natural language

explanation, contextual conversation, and pattern narration. The AI coach receives the

engine's real computed numbers as context and explains them — it never invents the

math itself.

This split mattered more than any single feature. It's the difference between a project

that *looks* AI-powered and one that *is* genuinely intelligent.

**Gemini API** powers the conversational coach and weekly AI-generated summaries — always

fed real context, never guessing.

**Firebase Firestore** stores monthly profile snapshots, enabling a deterministic

habit-change detector that flags "+17% this month, likely from increased food delivery"

— again, computed, not hallucinated.

**Deployment**: Initially deployed on Vercel for rapid iteration, with a Dockerized Google Cloud Run architecture staging for final production.

Because the carbon engine is pure functions, it was trivial to write 20+ meaningful

unit tests — ordering checks (vegan footprint < high-meat footprint), boundary checks

(risk score clamped 0-100), and invariant checks (Carbon Twin always projects lower

than current). This is testing that actually validates logic, not just "does it render."

Six structured master prompts: scaffold + security → deterministic engine → Firebase/AI

layer → core UI → coach/gamification UI → tests/Docker/README. Each prompt built on

the last, intent-driven the whole way.

🔗 Live: [https://ecosphere-ai-tan.vercel.app/](https://ecosphere-ai-tan.vercel.app/)

💻 GitHub: [https://github.com/ajx1tech/ecosphere-ai](https://github.com/ajx1tech/ecosphere-ai)