Persustain

Overview

Persustain is an AI-supported sustainability platform designed to help users better understand the environmental impact of their daily behaviors through explainable carbon estimation, personalized sustainability insights, and contribution-based tracking.

The project explored a central question:

How can AI systems communicate environmental impact in ways that are understandable, trustworthy, and meaningful to non-expert users?

My primary contribution focused on designing and implementing the backend AI architecture powering:

• Emissions estimation
• Activity normalization
• Sustainability insight generation
• Contribution aggregation
• Recommendation systems

Rather than using AI purely for prediction or automation, the platform treated AI as an interpretive layer, helping users understand how sustainability data is generated, what assumptions are being made, and how everyday actions connect to larger environmental outcomes.

Problem

Many green apps bury methodology or mix generative guesses with totals. The product goal was credible numbers plus readable explanations, not the other way around.

• Keep calculation paths deterministic
• Isolate AI for narrative and caveats only
• Scale cleanly as more users aggregate contributions

My role

Hands-on backend and AI implementation for the MVP.

• Cloud Functions architecture and APIs
• OpenAI and Climatiq integration
• Emissions and avoided-emissions semantics
• Firestore queries, aggregation, and cache keys

System design

Architecture split deterministic environmental math from generative interpretation: models never fabricated core emission totals.

That boundary improved reliability and made dashboards easier to reason about.

Emissions model

Early logic treated every logged action as additive emissions until we introduced avoided emissions and net framing.

• Actual CO₂
• Avoided CO₂
• Net impact

Virtuous behaviors (e.g. biking vs driving) showed as reductions versus baseline instead of oddly positive-only totals.

Product surface

What the backend backed in the MVP:

Technical development

Backend stack:

• Firebase Cloud Functions (2nd Gen)
• Firestore
• TypeScript
• OpenAI API
• Climatiq API

Representative callable units:

• normalizeAndExplainActivity
• generateBaselineInsight
• generatePatternInsights
• generateEvidencePackSummary
• recommendProjectAndGoals

Multi-user aggregation exposed stale cross-user reads when caches and keys weren’t fully user-scoped. Tightened query shapes, cache keys, and refresh behavior so aggregates stayed isolated and predictable.

Outcome

The final MVP included:

• onboarding-based emissions baselines
• AI-generated sustainability insights
• activity normalization infrastructure
• contribution aggregation systems
• recommendation logic
• scalable backend AI pipelines
• user-scoped dashboard architecture

The project was later discussed in the context of future:

• verification systems
• sustainability registries
• certification workflows
• scalable contribution infrastructure