Full-stack engineer · Paraná, Brazil

I build the systems small companies run their entire operation on —
multi-tenant CRMs, desktop ERPs, and AI agents that talk to real customers.
Four of them are in production with paying clients right now.

📫 Open to freelance projects and contract work · 🇧🇷 Atendo em português e inglês

Real systems with real users — three reachable on the open web, one installed on a client's machines.

Most of the code below is private client work. What I can show, I show.

Tracks a customer through 17 chained stages — lead intake → pre-sale → contract → finance → measurement → manufacturing → delivery → scheduled review — with per-tenant isolation enforced in the database, not in the app layer.

Next.js (App Router) TypeScript Supabase Postgres + RLS Zustand Vitest Playwright

The hard part — deploying to a machine a business is actively using. The deploy script always ships to the inactive color and never touches the live one. The running version stays up as both the rollback target and the drain path for in-flight connections; nginx flips only after the new color passes a health check. Images are built in CI and pulled from GHCR — nothing is ever built on the VPS, so a bad npm install can't take the site down.

57 numbered SQL migrations · 63 test files · 216 commits · e2e runs against an isolated tenant

An AI agent ("Marina") handles inbound real-estate leads over WhatsApp. A lead only becomes a card on the broker's Kanban once qualification actually closes — the model never decides that on its own.

NestJS Prisma PostgreSQL 16 Angular (standalone, signals) Redis MinIO OpenAI

The hard part — an LLM talking to paying customers can't be "mostly right". The AI service is split into entry-context → prompt-builder → generator → guardrails, with deterministic gates driving state transitions instead of the model. Prompt changes go through an offline eval harness: scenario fixtures (impatient lead, budget mismatch, discount negotiator, empty catalog, human-handoff request), a written rubric, and dated rounds with recorded findings. A persona change is measured before it ships, not eyeballed after.

Three services, ~81k LOC · 65 migrations · separate RAG service for catalog retrieval · Docker Compose + nginx on VPS

flowchart TD
  W[WhatsApp lead<br/>uazapi webhook] --> G[message grouper]
  G --> E[entry-context<br/>normalizer]
  E --> P[prompt-builder]
  RAG[(RAG service<br/>property catalog)] -.-> P
  P --> A[generator<br/>OpenAI]
  A --> R[guardrails]
  R -->|blocked| H[human handoff]
  R -->|passed| S{qualification<br/>gates}
  S -->|closed| K[Kanban card in the CRM]
  S -->|still open| W

Replaced the spreadsheets a freight company was running on. Windows desktop app plus a WhatsApp bot sharing one database. Modules: fiscal, financial, logistics, operations, registry, admin.

Electron React 19 Vite Supabase Sentry BullMQ NSIS auto-update

The hard part — desktop software you can't walk over and fix. Auto-update means a broken release installs itself on the client's machine. So CI gates every release behind tsc --noEmit and the full Vitest suite, and 10 Playwright e2e specs drive the real built app through login, base registrations, financial status math, logistics scheduling, quote→order, load scheduling and maintenance work orders. Brazilian NF-e invoice XML is parsed directly, so fiscal data is ingested rather than retyped.

105 test files · 52 migrations · 174 commits · shipping at v0.4.1 · public release channel

CLI that detects cloaking: landing pages that show one thing to an ad platform's reviewer and something else to real traffic.

TypeScript Playwright pixelmatch sharp Vitest

The hard part — you cannot detect a cloaker by fetching a page once. The tool fetches the same URL as two identities (bot and victim) through different proxies, follows the full redirect chain, and compares the results four ways: DOM structural similarity, extracted feature signals, parked-page heuristics, and pixel-level screenshot diff. Tests run against a local fixture server that deliberately cloaks plus a forward-proxy fixture — so the detector is validated against a controlled adversary instead of whatever the live web served that day.

Every run produces an evidence bundle: both HTML captures, both screenshots, the egress IP each identity exited from, and a timestamped trace of every request, redirect, cookie and console error.

VPSMAP — Every Docker UI shows you containers. Most of a real server isn't containers. VPSMAP merges Docker, systemd, listening ports and nginx vhosts into one live topology graph — and surfaces the routes that point at nothing. One Python file, standard library only, no agent and no database.

MeetScribe — Chrome extension that transcribes Google Meet by reading the native captions out of the DOM; no audio is ever recorded. Content script → service worker over an outbox that retries and only clears on ack, so a reload mid-meeting doesn't lose the transcript. Local IndexedDB history, TXT/Markdown export, optional AI summaries with your own API key.

• Migrations are the schema. Numbered SQL files committed to the repo — never a console click.
• Tests exercise the real thing. Playwright drives the built app, not a mock; the AI service is scored against fixtures with a written rubric.
• Deploys are reversible before they're clever. Blue-green with the previous version still warm beats a fast deploy you can't undo.
• Secrets come from env vars or a secrets manager. Never in the repo, never in a log line.
• Decisions get written down. ADRs when the choice was contested, PRDs when the scope was.

Also fluent in: Prisma · SQLAlchemy · TypeORM · Tailwind · Vitest · Playwright · Pytest ·
Supabase RLS & Edge Functions · blue-green deploys on VPS · OpenAI & Anthropic APIs · RAG · offline eval harnesses

Technology in Systems Analysis and Development Universidade Tecnológica Federal do Paraná (UTFPR) — Ponta Grossa

📫 pablotadinidev@gmail.com · LinkedIn