Link Charts — end-to-end engineering in production

Link Charts (linkcharts.com.br) is a URL shortener with advanced analytics, in production, that I created and maintain 100% solo. Every click is enriched with geographic, device, temporal and traffic-quality data, shown across 5 dashboards: overview, geographic (choropleth/heatmap), temporal, audience and insights.

In production: authenticated and public shortening, high-performance redirect with Open Graph previews for bots (WhatsApp/Telegram), anti-fraud with a per-click quality score, custom subdomains + link-in-bio page, public API with API keys, QR codes, tags, reports, CSV export, UTM builder, link passwords, expiration/scheduling/click limits, link health checks, retention e-mails and monetization via AdSense + Google Ads.

~1,929 commits across 3 repositories (725 backend, 1,029 frontend, the rest in docs), all mine, from Mar 2025 to Aug 2026 — kept up in concentrated bursts, alongside a full-time job. 50 release tags with independent semver per repository (backend at v2.16.0, frontend at v1.19.0).

Backend: Laravel 12 / PHP 8.2, PostgreSQL 15, Redis 7. Controller → Service → Repository layers with interface-based dependency injection, DTOs and ADRs. The critical /r/{slug} route serves HTML with Open Graph for bots and a 302 redirect for humans; tracking is 100% asynchronous via an idempotent job — no distributed lock: a dedup_key column with a UNIQUE index + insertOrIgnore guarantees a retry never duplicates a click. 10-min link cache; click enrichment in 3 phases (headers → server-side intelligence with viral rank/holidays → anti-fraud quality score).

Frontend: Next.js 15 (App Router) / React 19 / strict TypeScript, MUI 6 with an in-house design system, TanStack Query 5, ISR with cache tags and on-demand revalidation, en/pt-BR i18n, Auth0, CSP/HSTS in middleware, full SEO (JSON-LD, sitemap, llms.txt), 30 ApexCharts chart components and Leaflet maps.

Integration: proxy via rewrites (zero CORS), JWT in an httpOnly cookie (never localStorage), X-Request-Id propagated from the browser all the way to the queue worker for end-to-end log correlation. An in-house SQL dialect layer (SqlDateExpr) centralizes driver-dependent fragments so the test suite runs identically on SQLite and PostgreSQL.

OpenTelemetry with tail sampling (100% of errors + 100% of slow requests + 10% of the rest) exports traces, metrics and logs through Grafana Alloy — configured as code — to Grafana Cloud. Every trace carries service_version with the deploy SHA: a regression points straight at the release that introduced it.

8 per-domain log channels (redirect, tracking, jobs, auth, http, audit...) with automatic PII redaction and a request_id propagated from the browser to the queue worker. Faro RUM on the frontend mirrors the same build SHA; continuous profiling with Pyroscope/Excimer surfaces hot paths in PHP. 4 dashboards and 9 alert rules versioned as JSON in the repository — nothing hand-configured in the UI.

An external uptime probe runs every 5 minutes outside my own infrastructure and automatically opens a deduplicated incident issue when the service goes down — the safety net for the total outage that internal alerts, by definition, would never see.

Jul 13, 2026: 918 HTTP 502 responses during a deploy under the old model (merge published, build ran on the server itself) — including a real visitor coming from Facebook. That was the trigger to rewrite releases as tag-driven blue/green: warm up the new color, health-check it in a loop, graceful nginx cutover, drain the old color, abort automatically on failure.

A missing build-arg in the Dockerfile compiled the Google Ads conversion labels as empty strings — campaigns ran for weeks spending real money without registering a single conversion, invisible in dev because local builds read .env normally. The answer became a gate: check-build-args.sh compares every NEXT_PUBLIC_* referenced in code against the Dockerfile's ARGs and fails CI if any is missing.

The client IP was spoofable in logs and rate limiters; I fixed it with Cloudflare's real-ip and an automated test (ClientIpSpoofingTest) that keeps the flaw from slipping back in unnoticed.

The best evidence that blue/green works came from a deploy that failed: frontend v1.0.0 broke mid-pipeline (an rsync bug) while an external probe hit the site every 2s — 156 out of 156 samples answered 200. A broken release takes nothing down; it simply never happens.

I built Link Charts with a spec-driven, AI-first workflow: brainstorm → design doc → plan → execution, orchestrating multiple agents and subagents in phased runs with a report at every step.

Context as an artifact: a 22KB agent-context file (architecture guide) versioned in the repository guides the agents; ADRs and postmortems feed that context back over time; the frontend publishes llms.txt for LLM-readable content.

Automation with a human in the loop: an in-house command (/ship) automates commit → PR → CI → merge → deploy → health check, with at most 2 self-correction attempts per step — if that doesn't resolve it, it stops with an explicit warning and hands control back to me. The safe-migration rule also stopped being a wiki page and became a test that fails CI.

My thesis: solo dev + AI + strict gates produces company-grade output. The ~1,929 solo commits over 17 months, alongside a full-time job, didn't come at quality's expense: the same 902 tests, PHPStan and zero-warning checks blocked merges the whole time. AI amplifies; the guardrails decide.