stack

Languages with production delivery or a finished project behind them.

• TypeScript4 years

  Stech

  TaskForge, StechClub, PortalAZ and 16 more

• JavaScript4 years

  Outlier · Make

  Bubble Box Laundry, Apollo Digital, Braseng Business Consulting and 8 more

• Python4 years

  Stech · Cebrac · Outlier

  MITRA CRM, Mapping a collateral data pipeline, KitBuilder / S.A.G.A. and 3 more

  Version and dependencies upgraded in production financial systems, with the review closed on a clean pip-audit.

• Java2 years

  Pokémon in JavaFX

• C#2 years

  Persys

  AsuraPadel, BuscaJogos+, Schedule I — LocalMultiplayer

• PHP2 years

  Stech · Make

  Bubble Box Laundry, Braseng Business Consulting, Bumbum de Anjo and 5 more

• SQL5 years

  Sintel · Itaú

  Mapping a collateral data pipeline

• Go

  Portfolio API, Pronto, Brechoapp and 1 more

• C

  Sudoku Solver

What I use to put language models into a product with access control, evaluation, execution limits and measured cost.

• RAG with authorization before the prompt

  Restricted content is filtered at retrieval, and a negative test proves it never reaches the model.

• Hybrid search (BM25 + embeddings)

  In Jarbas, search combines BM25 (SQLite FTS5) and vectors fused by RRF, and the answer cites the file and line it came from.

• LLM evaluation and golden set

  Outlier

  Evaluated voice, vision and text models at Outlier, and keep a golden set to measure retrieval before swapping the engine.

• Review by a model other than the author

  In Conn-Pod one model executes and another verifies; in the real smoke test the verifier closed with status verified.

• Agents and orchestration

  TaskForge turns a request into a task graph and runs agents in parallel waves, one Git worktree per task.

• Action gate with human approval

  In TaskForge nothing runs without approval and a person does the merge; in Jarbas every agent write goes through a fail-closed gate that validates the plan first.

• Fan-out with concurrency cap and timeout

  In Jarbas, an agent's tasks form a graph: independent ones run in parallel up to the cap, each branch has a timeout, and a partial failure only counts as success when the contract says so.

• Durable queue with retry and dead-letter

  In Jarbas, agent work becomes a command persisted in SQLite with leases, retries, dead-letter and dedupe; in the test with a killed worker, the second attempt succeeded.

• Token cost per run

  I measure cost per model from real usage, with prices kept as versioned data and the cost left blank when a price is missing, never estimated.

• Agent isolation (worktree and deny list)

  In Conn-Pod each task runs in its own worktree under a sandbox, and the verifier got EPERM when it tried to read HOME, credentials and the network.

• Schema-validated agent output

  In Jarbas, skills and agents are versioned Zod contracts validated at registration, and every run returns a typed envelope with status, references, model and error.

• Anthropic, OpenAI-compatible APIs and CLI

  In Jarbas, the model sits behind a port with three backends: the Anthropic SDK, an OpenAI-compatible API and Claude Code in headless mode.

• Ollama, LM Studio and quantized local models on GPU

  I set up an internal server running Ollama in Docker on a 12 GB RTX 3060, serving quantized qwen3:8b; on my own machine I use Ollama and LM Studio.

• Prompt engineering

  Outlier

• Claude Code

  TaskForge, Conn-Pod

• React4 years

  Stech

  StechClub, PortalAZ, Professor Jarbas Financial Assessment and 1 more

• Next.js3 years

  Portfolio v2, Zoologic, Pronto and 3 more

• Angular2 years

  Persys

  AsuraPadel, Portfolio v1, Repository Explorer

• Vue

  BuscaJogos+

• React Native / Expo

  Pronto, Brechoapp, Pasta la Vista and 2 more

• Tailwind CSS

  Portfolio v2, Sekiro visual system

• HTML & CSS

  Make

  Bubble Box Laundry, Apollo Digital, Braseng Business Consulting and 9 more

• WordPress

  Make

  Viotto Clinic, Souza Lima - Music Courses, Souza Lima Music College and 1 more

• Core Web Vitals & technical SEO

  Acceptance criteria across 200+ landing pages in the Make cycle.

• Node.js3 years

  TaskForge, Conn-Pod, Jarbas and 1 more

• NestJS

  Stech

  StechClub, Pronto, OmegaHub

• Express

  Pasta la Vista

• Spring Boot2 years
• ASP.NET Core2 years

  Persys

• Django

  Mapping a collateral data pipeline

• Laravel / Lumen

  Stech

• PostgreSQL

  Stech

  Pronto, Brechoapp

• MySQL

  Stech

  StechClub, Optec Integrated IT Solutions, Zoologic and 1 more

• SQL Server

  Persys

  AsuraPadel

• Redis

  OmegaHub, Brechoapp, FinanceOS

• SQLite

  MITRA CRM, Portfolio API, Pokédex

• Prisma

  Stech

  StechClub, FinanceOS

• Indexes and execution plans

  Seven indexes derived from measured predicates took a screen from 4,271 ms to 354 ms.

• Kafka

  Mapping a collateral data pipeline, Pronto

• Webhooks & n8n

  PortalAZ

• Docker3 years

  Stech · Persys

  StechClub, Zoologic, BuscaJogos+ and 1 more

• GitHub Actions

  Portfolio API, Repository Explorer, Atalaia

• CircleCI

  Stech

• AWS S3 / EC2 / SES / SQS

  Stech

  StechClub, PortalAZ, Mapping a collateral data pipeline

• Azure App Service

  Persys

  AsuraPadel

• Vercel

  Portfolio v1

• Jest

  Persys

  AsuraPadel

• pytest

  KitBuilder / S.A.G.A., YoutCatcher

• JUnit

  Persys

• PHPUnit

  Zoologic

• Application security

  Audit and remediation across 9 financial-system repositories, CI pipeline included.

• Monorepo (Turborepo, pnpm)

  Portfolio v2, OmegaHub, FinanceOS

• ITIL & COBIT

  Sintel

• Power BI2 years

  Cebrac

• Power Query & DAX

  Cebrac

• Excel & VBA

  Cebrac · Itaú

• PySpark

  Mapping a collateral data pipeline

• Splunk

  Itaú

• Hadoop

  Itaú