EcoMitram
Community Climate Action Platform
EcoMitram ("Eco Friend") is a community climate-action platform serving India's national environmental movement, run by Paryavaran Sanrakshan Gatividhi (PSG) — quizzes, competitions, campaigns and certificates, distributed through NGOs and schools. 1.37M installs, 1.2M registered users (Play Store + Firebase console, September 2024).
We have engineered it as a volunteer project since 2020. Nobody pays for this work. It ships anyway — which tells you more about how we treat paid work than any brochure could.
Four generations, one commitment
| Gen | Years | What |
|---|---|---|
| 0 | 2020 | Adopted a volunteer-built "Earth Day" project (Laravel) and made it run |
| 1 | 2020–21 | Rebuilt with Laravel + Vue, shipped the first mobile app |
| 2 | 2022–25 | Full rewrite: Vue 2 + Express + MongoDB — 1,731 commits, three years in production |
| 3 | 2026 | Full rewrite again: Next.js 15 + NestJS — 600+ commits in five months, AI-assisted |
The scaling story
In one two-month campaign window (Jul–Aug 2024): 3 million visitors, 6+ TB of traffic, 804 million API hits — served from a single small VPS behind Cloudflare's free tier, with peak days of 86,000 users. Zero marginal infrastructure spend, by design: a volunteer project earns its scale through architecture, not budget.
Same era: a national quiz produced 225,000 free-text school-name entries — messy, duplicated, inconsistent. A fuzzy-matching pipeline (published as open source) merged them to 74,000 clean entries. (Classic string-similarity automation — we don't call it AI, because it wasn't.)
The 2026 rewrite — where AI actually earned its keep
The v3 rebuild is the most AI-intensive delivery we've run:
- 600+ commits in five months, single human operator, with heavy Claude co-authorship across multiple model tiers — at peak, ~90 parallel worktrees and 200+ agent sessions in a single month
- The repo is built to be operated by agents: five CLAUDE.md files, 295 markdown docs, and CI that fails the build if the auth-policy doc drifts from generated reality
- AI in the product, tested and shipped: OpenAI-based content moderation for community posts; an AI eco-tagger classifying the affiliate product catalog
- Security work with receipts: a privilege-escalation vulnerability found and fixed with a regression test, plus a formal phased security/quality audit (P0–P5) run across the codebase
- Accessibility as standard: WCAG AA internal guideline, aria coverage across 600+ files, reduced-motion support — CI-enforced, not aspirational
What we tried and dropped — because honesty is the point: an AI-prompt-based certificate importer lost to a plain drag-and-drop editor; the AI version was less reliable. A real org hit "Campaign not found" trying to self-serve a campaign; we redesigned the model around admin-created campaigns instead of patching the error. Real users beat clever features.
Where it stands
Active daily development. The platform serves PSG's national campaigns — tree plantations, school competitions, eco-pledges — through orgs that get shareable campaign links, QR codes, and participant tracking.