Software Engineer · Backend & AI Integration — Recife, Brazil · open to remote and relocation

I build backend systems and put LLMs to work inside real products: retrieval, agents, evaluation and the guardrails that keep them honest. Currently a Software Engineer at Coopanest-PE, working on the company ERP (PHP/Laravel/MySQL), where I shipped an OCR pipeline with the Gemini API that automates billing documents. Postgraduate student in Deep Learning; B.Sc. in Computer Science (CESAR, Recife).

Stack   Python · FastAPI · PHP/Laravel · TypeScript/Node.js · PostgreSQL + pgvector · Supabase · MongoDB · Docker · AWS · Terraform · LLMs (OpenAI, Anthropic, Gemini) · LangChain · LangSmith

FastAPI + Supabase backend and React frontend for finding and joining local events.

• LLM "taste agent" turns free-text answers and event descriptions into pt-BR keywords; pgvector embeddings match people to events, with a deterministic fallback when the model fails
• Waitlist with "call next", JWT-signed QR check-in, trophies, moderation, rate limiting (Upstash Redis), Postgres RLS
• Apify ETL for the event catalog, LangSmith tracing, deployed as serverless on Vercel
• ~1,050 automated tests (669 pytest + 381 Vitest) · frontend repo

Monorepo (Next.js, TypeScript) that listens to a live broadcast, extracts checkable claims and publishes short alerts — always with a source.

• Pipeline: streaming STT (Deepgram) → claim extraction → hybrid retrieval over fact-checks, official indicators and documents (pgvector + full-text, RRF) with a penalized web fallback → verification → scoring
• Human-in-the-loop editor approval; DB-level guarantee that a correction without a verified source can't be published
• Labeled eval set measuring precision, false alarms and p50/p95 latency; threshold sweep without calling models

Deterministic lead qualification and scheduling over WhatsApp (Twilio, Meta Cloud API, Z-API), built as the measurable baseline for a future AI agent.

• Pure state machine (handle() does no I/O, enforced by tests) with an engine that executes effects; every reply comes from a reviewed message catalog
• Multi-tenant Postgres with tenant isolation in the schema, double-booking prevented by row locks plus a partial unique index, outbox delivery to a Laravel CRM
• 456 tests (pytest)

Infrastructure as code for CodePipeline → CodeBuild → ECR → EKS, building and deploying a containerized app on every push.

Turns Deep Learning lecture transcripts into summaries, active-recall flashcards and a podcast per class: 4h of lecture into 20 min of reading.

LinkedIn · warpedro1@gmail.com · English (full professional) · Portuguese (native)