AI Engineer · Brazil
I build AI systems that run in production, not demos: LLM agents that handle 300+ leads a day, the server-side tracking that proves those leads converted, and the stack underneath — Postgres schemas, webhooks, auth and the front end.
Most of what I ship day to day is private code for enterprise clients, so this profile has two halves: the systems I built but can't open, and the work you can audit yourself.
AI SDR agents · WhatsApp and Instagram — qualification, meeting booking and direct CRM writes (Bitrix24, RD Station), handling 300+ leads a day. Meta delivers the same webhook two or more times in parallel, milliseconds apart, so deduplication has to be cross-execution rather than in-memory; conversations that leave the agent's scope are handed to a human instead of guessed at.
Server-side conversion tracking — Meta CAPI and Google Ads offline conversions fed from CRM stage changes, which brought back attribution for campaigns reporting zero conversions. The root cause is rarely the pixel: it is usually the funnel the form posts to versus the one the automation reads.
Multi-tenant SaaS CRM · 862 commits, 20 active clients — sole developer, end to end: PostgreSQL schema with row-level security and Edge Functions, WhatsApp Inbox on the official Cloud API, Kanban pipeline, dashboards, and a React/TypeScript/Vite front end with route-level code splitting.
Authentication and social login — Google, TikTok and Instagram on OIDC/OAuth 2.0, replacing Keycloak with an in-house JWT layer (access token plus rotating refresh in an httpOnly cookie) in a Next.js + MongoDB product.
RAG agents with layered memory — triage, FAQs and human handoff over an embedded knowledge base, cutting first-response time by 60%.
Six patches merged upstream so far, in Supabase and QuestDB:
- supabase/supabase-js: encode broadcast header fields as UTF-8 · match response Content-Type case-insensitively
- supabase/supabase: Apple Services ID must be first in Client IDs for web sign-in
- questdb/documentation: TTL on materialized views · query.timeout duration key · Parquet and CSV export in the Result Grid
The one I'd point at first is supabase-js#2516: a UTF-8 byte-length mismatch was silently corrupting Realtime broadcast headers whenever a channel name carried an accent or an emoji. One side counted characters, the other counted bytes, so every test passed in plain English. Fixed with a regression test that failed before and passed after.
Open in review: PKCE verifier, storage metadata encoding, postgrest filter escaping and rpc over POST in supabase-js; JSON.parse hardening across Studio; null guards in postgres-meta; CSV quoting and lib hardening in cal.diy; plus one each in Appsmith, ToolJet, Infisical and QuestDB docs.
- IdeiaTaker — meeting transcription and AI summaries for Google Meet with no bot in the call: it reads the native captions, so no extra participant joins and no audio is recorded. Chrome MV3, Next.js, Supabase, Groq and Gemini. In production.
- brain-template — persistent, git-versioned memory for AI coding agents: semantic facts, session handoffs and skills in plain Markdown, shared across Claude Code, Gemini CLI and local models. No vendor database, no auto-compaction.
- claude-dongle — usage-limit monitor for Claude Code: burn rate, overflow forecast and per-project consumption over the official OAuth usage API. Python, PyQt6.
- BodyWay — personal trainer app, Flutter + Supabase (RLS, 20 Edge Functions), live on Google Play and the App Store.
- Cromo Certo — freemium marketplace for World Cup 2026 sticker collectors: trades, auctions and on-device album scanning. React 19, Supabase, Turborepo.
Based in Brazil (UTC-3), working remotely with teams in the US and Europe. English: professional working proficiency.