Open to remote engineering roles (LATAM / global) · GMT-3, overlap-friendly with US East / EU West · Santa Catarina, Brazil

I build backend-heavy systems that ship to production: AI agent orchestration, full-stack SaaS, and native macOS. Depth in Python backend and LLM agent systems; breadth across the stack — from a LaTeX-rendering PDF pipeline down to a reverse-engineered HID protocol. The projects below are ordered by how directly they map to a hiring need, not by recency.

Reach me: LinkedIn

• quarterdeck — multi-agent orchestration for Claude Code
• forja — CV-builder SaaS, live preview + LaTeX → PDF
• local-mind — local-first memory for AI agents

An orchestration framework for AI coding agents: 26 agents across 8 squads, executed in parallel waves.

The single-agent default serializes multi-step engineering work into one bottleneck. Quarterdeck decomposes a request into waves — within a wave, 3–5 specialized agents run in parallel (reconnaissance, validation); waves run sequentially only when there's a true data dependency. The design rests on a few deliberate constraints:

• Hierarchy as an invariant — agents never act on their own; an orchestrator dispatches and synthesizes. This keeps a multi-agent system debuggable instead of emergent.
• Zone assignment before parallel writes — write-capable agents get explicit file ownership so no two touch the same file in one wave. Read-only reviewers (quality gates) skip this and always parallelize.
• Deterministic routing + standardized output — requests route to squads by signal; every agent returns the same output contract, so synthesis doesn't depend on per-agent formatting.

Python · multi-agent systems · parallel orchestration · routing

A CV-builder SaaS with live preview and LaTeX → PDF export — no local toolchain, no vendor lock-in.

A full-stack product running in production. The engineering load sits in three places:

• LaTeX PDF pipeline under concurrency control — PDFs compile with Tectonic (XeTeX) from EJS → LaTeX templates; a p-queue with configurable PDF_CONCURRENCY caps server load, and LaTeX-character escaping closes off template injection.
• End-to-end auth & security — Better Auth with DB-backed sessions tracking IP and User-Agent, user/admin roles, per-route fixed-window rate limiting, httpOnly cookies, Zod validation on every endpoint, and parameterized queries through Drizzle ORM.
• Applied AI — CV parsing, ATS scoring, and translation via Groq / Llama 3.3 70B, with a 24h in-memory TTL cache on ATS scores to cut redundant API calls.

Hono · React 19 · Vite · Better Auth · Drizzle ORM · PostgreSQL · Tectonic

Local-first persistent memory for AI coding agents — SQLite, zero cloud, zero API keys.

A clean-room rebuild: same goal as a cloud-based predecessor, opposite architecture — local SQLite instead of a cloud API.

• Custom relevance scoring — ranking by BM25 * 0.7 + Recency * 0.3, with recency decaying as e^(-0.15t) (≈4.6-day half-life). Severity tunes the decay: CRITICAL findings persist ~35 days, INFO decays in ~3.5.
• Data engineering — a 9-table SQLite schema with FTS5 full-text search and incremental transcript capture by byte-offset (only new bytes are read).
• Test rigor — 18 suites / 310 tests, Biome for lint/format, esbuild for builds.

JavaScript · SQLite · FTS5 · information retrieval

These ship the same product instincts down to the protocol and FFI layer.

A native macOS SIP softphone — "the MicroSIP of the Mac." Engine: PJSIP 2.17.

Low-level systems work across a three-language boundary. Three decisions carried it:

• A clean FFI boundary — PJSUA2 (C++) is subclassed in Objective-C++ and exposed to Swift as a pure Obj-C delegate of immutable value objects, so C++ never leaks above the bridge.
• Threading discipline over GCD + pjlib — one serial dispatch queue owns every PJSIP call; since GCD doesn't pin threads, each block re-registers via pj_thread_register before touching pjlib — the fix for intermittent asserts that only reproduce under load.
• Testable against a frozen protocol — the whole UI runs against a SIPEngine protocol; a FakeSIPEngine (ORELHAO_FAKE_ENGINE=1) made the UI buildable and demoable before the real engine existed.

The README's hard-won gotchas go deeper: port 5060 self-answering, INVITEs over 1300 bytes silently dropping to TCP per RFC 3261 §18.1.1, UDP undelivered under Docker Desktop with network_mode: host.

Swift · Objective-C++ · SwiftUI · PJSIP · SIP / RTP

A macOS configurator for the Attack Shark R1 mouse, built on a reverse-engineered HID protocol.

The vendor ships Windows-only software; the protocol was reversed and documented from scratch (docs/PROTOCOL.md).

• HID reverse engineering — a vendor HID interface (usage page 1 / usage 0x80, 64-byte feature reports); button remaps write via feature report 0x08, firmware ACKs on 0x03. Talks to the device directly over IOKit — no third-party software at runtime.
• Layered design + golden-vector tests — R1Kit holds protocol + IOKit transport with no UI (ReportBuilder, HidTransport, ScrollInverter); CLI and SwiftUI app consume the kit; report encoding is locked down by golden-vector codec tests.
• An OS-level workaround for a firmware gap — the firmware has no native scroll-invert, so per-device inversion runs macOS-side through a CGEventTap login agent, leaving trackpad natural scrolling untouched.

Swift · IOKit · HID protocol · reverse engineering

Backend & AI — Python · FastAPI · LLM agents (Claude Code) · Groq / Llama Web — TypeScript · Hono · React · Vite · Node · Drizzle ORM Systems — Swift · Objective-C++ · IOKit · SwiftUI Data — PostgreSQL · SQLite · FTS5

Open to remote engineering roles, LATAM or global — backend-heavy or full-stack. The fastest way to reach me is LinkedIn