I want to share a slightly unusual Showcase post: the output isn’t an app or a game — it’s an entire company. And I’m not an engineer.
I’m Artem, founder of HERD — an open-science project building a dense network of $25 infrasound sensors (MEMS barometers + ESP32 + GPS timing) for early warning of natural hazards: earthquakes, volcanic eruptions, events like the 2022 Tonga Lamb wave that reached barometers hours ahead of the water. It’s named after the elephants that moved uphill before the 2004 tsunami hit Phuket, where I live. To be scientifically honest: we detect the infrasound of the source event — not the tsunami wave itself.
I don’t write code. Everything below was built by me directing an AI agent inside Cursor, over a matter of weeks:
- A 6-language website (EN/RU/TH/ZH/JA/ES) — 255 HTML pages, 160 public URLs, deployed to Cloudflare Pages by the agent via wrangler from the built-in terminal.
- A science library: 13 long-form articles × 6 languages with 272 vetted sources, deduplicated by DOI. When the count grew from 159 to 272, the agent updated it consistently across 32 places on the site in one session.
- Company infrastructure: Delaware LLC via Stripe Atlas, Mercury banking, live Stripe payments — the agent drafted the applications, wired the payment links into every language mirror, and checks the account state through Stripe’s MCP integration.
- A content machine: 22 Python scripts I could never have written — a video-render pipeline for Instagram/YouTube reels (PIL + ffmpeg), a fact-card generator, an IG auto-publisher with a cron scheduler and a 96-post queue, WebP conversion, sitemap generation.
- An internal “mission control”: a password-protected
/opsdashboard on Cloudflare Pages Functions (Basic Auth middleware + a read-only Stripe stats endpoint) so I see status and money in one place.
What actually made this work in Cursor (the part I’d recommend to any non-developer):
-
Project rules as a “bible”. Three files in
.cursor/rules/: a context primer (what HERD is, what the science honestly can and cannot do), a deployment rule (exact deploy command, what never to break), and a dashboard rule. Every session starts with the agent already knowing the project better than a new hire would after a week. -
A living dashboard file as single source of truth.
PROJECT-DASHBOARD.md— the agent reads it at session start and updates it when status changes. This killed the “the AI forgot what we decided last week” problem completely. -
Multi-file agent edits across language mirrors. The same change applied to 5-6 locale versions of a page in one prompt — this is what makes a 6-language site maintainable for one person.
-
Anti-drift discipline written into the rules: when a load-bearing fact changes (an engineer review corrected our sensor architecture, a tax lawyer corrected a claim), the rule says: update the bible and dashboard first, then propagate to pages. The agent enforces the order.
-
Honest status: this is pre-prototype. The first physical sensor is designed (spec’d down to the BOM — ~$14 in parts at first-batch quantities) but not yet built; ordering the first parts is the current top priority. The promo film has ~100k views, payments and lead capture are live, and the first sponsor negotiation is underway. The biggest open question is the science validation — and that’s exactly what the first sensor is for.
-
I used to think “AI coding tools” were for developers to go faster. My experience is different: Cursor let someone with domain knowledge and zero engineering background do the work of a small team — site, science, legal, payments, content ops. The bottleneck stopped being “can I build it” and became “what should exist.”
-
Happy to answer anything about the workflow (rules setup, the dashboard pattern, how I QA what the agent ships when I can’t read the code).
