case study

i built my own editorial operating system

i was a software engineer. then i became a creator. in 2026, working with AI coding agents, the two finally met: one person built the data platform, the intelligence layer and the publishing automation that run a 250k-follower creator business — no team, no SaaS, on a $12/month server.

the numbers it runs on

live figures, straight from the system this page describes. last update: 15 jul 2026.

248kcommunity across instagram and tiktok (plus youtube, facebook and threads)

7.5maccumulated views on tiktok

45kaudience comments ingested and classified by the intelligence layer

~54python scripts, standard library only — the whole engine

5social networks measured hourly through official APIs

$12per month — one self-hosted server runs everything

the story, in three acts

act one

the software engineer

i spent my first career in tech as a software engineer, and more than 15 years of my life in london. building systems for other people's businesses taught me how data becomes decisions — and how much of any operation is just well-organised plumbing.

act two

the creator

i left that career to talk about what i actually care about: human behaviour, relationships, fatherhood, race — in portuguese, as an equal. the work grew into a community of 250k+ people and tens of thousands of real conversations in the comments. and with it came the creator's classic problem: five networks, scattered metrics, zero infrastructure — and no team.

act three

the editorial os

AI coding agents collapsed the distance between having an idea and having a system. working in plain language with claude code, i designed and shipped my own platform in weeks: pipelines, dashboards, an intelligence layer over my audience's comments, and a blog that publishes itself. the engineer i was, the creator i am — with AI in the middle.

what it does

measure

every network, one panel

hourly ETL from instagram, tiktok, youtube, facebook and threads via official APIs, plus site traffic from server logs and google search console. a single "panorama" shows community, reach, search and AI-presence in one place — with a morning report and a live TV wall.

understand

comment intelligence

45k audience comments ingested and classified — by deterministic rules, no black box — into themes, emotions, questions, pains and product signals, then crossed with performance data to decide what to make next.

create

data-driven editorial

an editorial planner that crosses what the audience searches for (keyword research) with what actually performed, generating scored content briefs — so the next video starts from evidence, not guesswork.

publish

a blog that ships itself

every video becomes a long-form article: drafted into headless wordpress, staged as a trello card, cover attached, published automatically and deployed as a static site — with entity-level SEO and a daily monitor measuring whether AI assistants cite the site.

under the hood

sources

instagram graph api · tiktok, youtube, facebook & threads (via windsor.ai) · nginx access logs · google search console · gemini (with search grounding, for the AI-presence monitor)

pipelines

~54 python scripts, standard library only — pullers on cron write JSON feeds and SQLite histories (audience demographics, comments, daily snapshots). no framework, no dependencies, nothing to break.

intelligence

rule-based comment classifier (themes · emotions · questions · pains · product signals) → anonymous aggregates only ever leave the server. theme × performance analysis ranks where to invest.

surfaces

static dashboards that fetch live JSON client-side, behind a custom auth service (pbkdf2 + nginx auth_request, per-page permissions) · a public TV wall · a morning report

publishing

headless wordpress ↔ trello automation → static site generator → denisonluz.com, with JSON-LD entity graph, llms.txt and hreflang

built with

AI coding agents (claude code) doing the building; me doing the deciding, reviewing and constraint-setting. the discipline that kept it maintainable: standard library only, generators over hand-edited files, secrets never leave the server.

the stack

python 3 · stdlib onlysqlitenginxdigitalocean ($12/mo) headless wordpresstrello apimeta graph apiwindsor.ai gemini apiclaude codezero node_modules

the point isn't the dashboard.

it's that one person, with AI agents and clear constraints, can now build the internal tooling that used to take a team. if you want to talk about this case — a talk, a workshop, a conversation — write me.

contact@denisonluz.com