FLORENTIN PURCEA
RO

Work

Systems I've built.

Not a portfolio of screenshots. Each entry says what hurt, what I built, on what, and what can be verified.

Systems I've built

in productionPerformance marketing agency — 12 client accounts

A multi-agent advertising operating system

Challenge

Twelve advertising accounts across Google, Meta and TikTok, each with its own tracking, budget and rules — and one person to watch all of them. Silent failures (delivery stopped, tracking on consent, geo leaks) were found weeks late.

System

One brain per account (281 governed tools), twelve agency-core agents (health, incidents, analysis, strategy, alerts, prompts, competitive and creative intelligence), daily sync into an owned database, a read-only 15-rule weekly check with evidence, Excel reporting to Drive, and a decision log where every change is proposed, approved, executed once, measured and reversible.

  • Cloudflare Workers
  • D1
  • Vectorize
  • Service Bindings
  • Google Ads API
  • Meta Marketing API
  • 24 workers live, 5 scheduled automations, 12 accounts synced daily
  • A lead event firing on cookie acceptance found and fixed: ~1,900 "leads" of which ~180 were real
  • Weekly check proven read-only by structural tests
stagingDental clinic, Bucharest

A dental clinic conversion system with an AI concierge

Challenge

A live site with LCP of 11.9 seconds, no booking flow, calls missed after hours and no view from ad to appointment.

System

40-page rebuild with a Treatment Navigator and symptom selector, a 3-step booking, an AI concierge with medical guardrails and a server-side emergency gate that can book in chat, a lead console with a missed-lead watchdog and repeat-caller detection, and a zero-loss SEO migration map.

  • Astro
  • Cloudflare Workers
  • D1
  • Workers AI
  • Playwright
  • Lighthouse 93 / 100 / 100 / 100 on mobile
  • 30 end-to-end tests green
  • Every legacy URL mapped before cutover
in productionTowing & roadside assistance, Bucharest — remorcari-tractari.ro

A phone-first emergency service platform, live

Challenge

A stressed person on the side of the road at night, one hand free, on a slow WordPress site with 207 URLs of SEO history that could not be lost.

System

Static-fast rebuild built for one-tap calling and WhatsApp, 10 service pages, 14 zone pages behind an anti-spam differentiation gate, 6 guides, schema and llms.txt, GA4 with Consent Mode v2, a validated redirect map and a plate-blur media pipeline. No forms, no database, no secrets — by the owner's decision.

  • Astro
  • Cloudflare Workers
  • GA4
  • Consent Mode v2
  • Playwright
  • Live in production since 2026-08-30
  • Lighthouse 100 / 100 / 100 / 100 mobile and desktop
  • 207 legacy URLs kept, redirected or retired — none lost
stagingAgency product, piloted on a dental clinic

An AI-search visibility operating system

Challenge

Nobody could say, with numbers, how AI assistants describe a brand or whether a content change moved anything.

System

A closed loop — discover, measure, diagnose, prioritize, fix, validate, learn, attribute — sampling AI engines through official APIs, every rate with a sample size and a Wilson interval, brand-truth facts checked against AI claims, opportunities turned into briefs and experiments, GA4 AI referrals attributed, weekly runs on a durable scheduler, client report archived.

  • Cloudflare Workers
  • D1
  • R2
  • Durable Objects
  • Search Console API
  • GA4
  • 13 MCP tools, a client portal and a weekly report on staging
  • Two approved opportunities executed on the pilot site
  • Nothing shown under n < 5 — by rule
stagingAccounting firm, Bucharest

A professional-firm site with a quote builder and a verified assistant

Challenge

Prospects wanted a price range now; juniors answered the same tax questions daily; the site was a brochure.

System

56 prerendered RO/EN pages, a QuoteBuilder (scope → indicative fee), an assistant that answers only from an approved knowledge base and never invents tax rules, a knowledge hub for SEO/GEO, a lead console with GDPR consent, and a demo kit for sales presentations.

  • Astro
  • Cloudflare Workers
  • D1
  • Workers AI
  • Vitest
  • Playwright
  • Lighthouse 100 / 100 / 100 / 100
  • 15 end-to-end tests green
  • Presentation-ready on a private staging URL
in productionConsumer health product, live

A lab-results explainer, GDPR art. 9 by design

Challenge

People google each value on their lab report in fear. Health data is the most sensitive category there is.

System

Upload the PDF; a deterministic parser reads every value and the interval printed on that report; the model only writes the words; every value cites its page and row; explicit health-data consent before anything; one-click export and erasure; one cookie, no third parties. Never diagnoses.

  • Next.js
  • Cloudflare Workers
  • D1
  • R2
  • Claude
  • Live in production
  • Verified end to end with real reports, trend across reports of the same person
  • Zero analytics, zero third-party requests
stagingHome-care e-commerce, ~1,500 orders/month

An edge storefront replacing WooCommerce

Challenge

WooCommerce at 1,500 orders a month: plugin conflicts, slow checkout, hosting bills, and e-invoicing that must keep working.

System

A custom storefront with the catalogue in an edge database, a back-office designed for the order volume, historic order import planned for 50–100k orders, invoicing and payment providers kept, tracking and consent built in.

  • Astro
  • Cloudflare Workers
  • D1
  • Playwright
  • 579 URLs on staging, accessibility 100, LCP under 2.5s
  • 509 products populated in the edge database
  • Production untouched until the client's go
in productionInternal operations

A control plane for a 51-project portfolio

Challenge

Fifty-one projects, sixty-five workers, nine Cloudflare accounts — and no single place that knew what was deployed, broken, stale or ready to launch.

System

Discovery across GitHub and Cloudflare, daily audits (performance, accessibility, screenshots) from the edge, health and readiness scores where UNKNOWN is never zero, launch gates with waivers, regression sweeps, and recommendations whose impact is measured after they ship.

  • Cloudflare Workers
  • D1
  • Browser Rendering
  • GitHub API
  • Google Sheets
  • MCP
  • 36 MCP tools, 51 projects registered, audits running daily
  • 19 recommendations validated with before/after numbers
  • Read-only toward every child repository — by rule

Custom systems

Don't see what you need? I can build it.

Tell me the problem, the people involved and what "done" looks like. I answer with an architecture, a complexity estimate and the first step — usually within a day.

Tell me what you want to buildWhatsApp Me
EmailWhatsApp