I build applied AI products end to end, and my main project right now is Mooose, an open-source ENEM essay grader: LLM orchestration, RAG pipelines, FastAPI services, and the observability that keeps them honest in production. I work at Dígitro Tecnologia, ship SaaS on the side, and study Information Systems at UFSC.
📍 Florianópolis, Brazil · 🎓 Information Systems @ UFSC (2024–2028) · 🗣️ Portuguese (native) · English (B2) · Spanish (B1)
Mooose is the open-source ENEM essay grader I build in public. ENEM is Brazil's national exam, and millions of students train essays with nobody to grade them. Mooose grades each of the five official competencies, explains how to improve, and runs on whatever LLM you bring: free OpenRouter models, Gemini, Groq, a local Ollama, OpenAI or Claude. The hosted version has graded 1,000+ essays.
What makes it interesting as an AI engineering project:
- LangGraph pipeline: theme analysis, five specialist graders in parallel, a reviewer that can move a score by at most one level, deterministic calibration, then synthesis.
- LLM + rules, on purpose: ENEM annulment rules, tangency caps and the Competency 5 score (computed from an agent/action/means/purpose checklist) are unit-tested Python, not model opinion.
- Self-consistency: sample each competency N times and take the median, which steadies free and small models.
- Provider-agnostic: bring-your-own-key API with SSE progress streaming; structured output that works without tool calling, so it runs on free and local models.
- Measured, not guessed: offline tests with a scripted fake LLM (29 backend + 9 frontend tests, CI on every push), an eval harness, and prompt-injection hardening.
- Photo input: handwriting read by a vision model (a free one worked), with local PaddleOCR as an option.
![]() |
![]() |
⭐ Star it · 🌐 Project page · 🧩 Good first issues · 💖 Sponsor
moodle-academic-mcp is an MCP server that connects a student's own Moodle to Claude Code, Codex and ChatGPT. Ask "what is due next week?", "how many classes can I still miss?" or "what do I need on the P2 to pass?" and get answers from your data, with your token and your model.
- 55 tools from one operation registry: the same registry serves MCP (stdio + Streamable HTTP) and REST/OpenAPI for GPT Actions.
- Exact math, not model guesses: attendance and grade simulation use
Fraction/Decimal; a grade not posted yet is never treated as zero. - Security first: tokens are redacted from logs and results, SSRF is blocked, and text from PDFs and forums is treated as untrusted data, never as instructions.
- Honest about validation: a public feature matrix says what ran on a real Moodle and what only ran against the mock.
⭐ Star it · 🧩 Good first issues · 🗺️ Roadmap · 🏫 Add your university
Self-hosted: both cards are generated daily by a workflow from the GitHub API. No third-party widget, nothing to go down.
Every pull request I send to a project that isn't mine becomes a star. Stars from the same project form a constellation. Green is merged, blue and pulsing is still open. Hover a star to see which PR it is.
Latest contributions (auto-updated)
Merged: 20 in 8 projects
In review: 11
No chatbot to sign up for. Click a question, press Submit new issue, and within seconds a tiny retrieval bot (BM25 over my resume, no LLM, no API keys, source) replies in the thread.
| Does he know RAG? | What did he build for telephony? |
| How does he evaluate LLMs? | What is Smart Pricing? |
| What is his tech stack? | Ask your own question → |
AI Developer — Dígitro Tecnologia · Feb 2026 – now · Florianópolis (on-site)
- Enterprise generative AI with LLMs and orchestration patterns (tools, routing, memory, fallback) to automate internal flows.
- End-to-end RAG: ingestion and normalization, chunking, embeddings, vector store and re-ranking, to raise precision and cut hallucinations in production.
- FastAPI services between LLMs and internal APIs, with schemas, versioning and error handling.
- Telephony bots: call-transfer flows and emergency-call classification, integrated with IVR systems.
- No-code-driven agents that generate IVR code and configuration (FastAPI + LangChain), cutting delivery time by about 2 days.
- Internal LLM evaluation framework (quality, cost and latency; models, embeddings and re-rankers) and observability (structured logs, per-request tracing, latency/cost metrics).
- Scrum Master for the team: prioritization, unblocking and stakeholder alignment.
Previously AI Developer Intern at Dígitro (Apr 2025 – Jan 2026): VoIP call diagnosis with LLMs, re-ranking evaluation pipeline, model endpoints and test datasets.
Full Stack & AI Developer — Smart Pricing · Nov 2025 – now · Remote (contract)
- Built and shipped a SaaS for vacation-rental hosts (Python/React) covering pricing, occupancy and performance for portfolios of 30+ properties.
- Multi-tenant FastAPI REST API, React dashboards, pricing automations (unbooked-day and occupancy-gap rules), and ADR / occupancy / REVPAR metrics.
- Alerts for negative reviews, scheduled metric jobs, and an AI proof of concept for automatic insights per property.
AI & Automations Developer — Fonte · Feb – Apr 2026 · Remote (contract)
- n8n workflow over the WhatsApp Cloud API, OpenAI and Trello that turns group messages into structured, validated tasks, with Redis duplicate control, automatic deadlines and error handling.
Nexxera · Implementation Operator (Dec 2023 – Apr 2025) · Implementation Apprentice (Jan 2022 – Dec 2023)
- Automated implementation routines and validations with Python and Shell Script; audited corporate data with SQL.
| Project | What it is | Stack |
|---|---|---|
| Mooose ⭐ | Open-source ENEM essay grader: LangGraph pipeline with five specialist graders, rule-based calibration, bring-your-own LLM (free options included). 1,000+ essays graded on the hosted version. | Python · FastAPI · LangGraph · LangChain · React · Docker |
| moodle-academic-mcp | MCP server that plugs a student's own Moodle into Claude Code, Codex and ChatGPT: deadlines, delivery with read-back evidence, exact attendance and grade math. Open to contributors. | Python · MCP · FastAPI · SQLite/PostgreSQL · Docker |
| Real-time face recognition | Local face detection and identification in real time. | Python · OpenCV · NumPy |
| Toxic content detector | Comment classification pipeline with a simple UI for testing. | Python · scikit-learn · pandas · Streamlit |
| Limpa CNPJ | Chrome extension with 170+ active users. | JavaScript · HTML · CSS |
AI: LLMs · RAG · embeddings · re-ranking · prompt engineering · model evaluation · observability | Data: PostgreSQL · pgvector · Qdrant · Redis · Oracle · MySQL | Ops: Docker · CI/CD · Grafana · Render · Vercel
🏅 Google AI Essentials · Databricks Generative AI Fundamentals · OBMEP honorable mention (2019, 2022) · ONC silver medal (2022)






