Rafael Pupio Vieira · AI Engineer

Rafael
Pupio
Vieira

Work authorization — no visa needed: US Brazil EU·27 + EEA Switzerland

US · Italian · Brazilian citizen — plus simplified Mercosur residency rights across South America

AI Engineer building agents & automation on 10+ years in administration, operations & financial operations

10+
Yrs admin · ops · financial ops
3
AI certifications
3
Citizenships
1
Trading agent — retrains daily
Builds Daily

The Builds

FOLIO 01

Live cargo — what I'm building, how the AI is actually applied, and what it runs on.

001

Daily-Learning Trading Agent

In Build

An end-to-end AI day-trading R&D pipeline. The core is a deep-learning agent architected to retrain on every trading session — each day's market data feeds the next day's model, so the system compounds instead of going stale. Claude Code operates as an active engineering partner across architecture, code, and research.

AI appliedDeep-learning agent design · daily retraining loop · agentic engineering with Claude Code driving architecture & implementation
Domain edgeMarket structure & risk framed by 10+ years of FP&A discipline — the model's objectives are set by someone who has owned a P&L forecast
StatusActive R&D — architecture documented, agent in iterative build
Deep LearningAgent ArchitectureRetraining PipelineClaude CodeAgentic EngineeringQuant Research
002

Persistent Context System

Shipped

The knowledge layer that makes the trading agent buildable: TRADING-KNOWLEDGE.md and DAILY-LEARNING-AGENT.md — foundational documentation engineered as durable LLM context. Dropped into the project repo, they brief every fresh Claude session on the domain, the architecture, and the decisions already made. No re-explaining, no drift.

AI appliedContext engineering · documentation-as-memory · repo-native knowledge base that turns a stateless model into a briefed teammate
Why it mattersAgentic projects live or die on context quality — this is the difference between an assistant and a collaborator
StatusShipped — in daily use powering build 001
Context EngineeringLLM MemoryKnowledge Base DesignPrompt EngineeringDocs-as-Code
003

This Site

Live

You're looking at build 003. Hand-rolled HTML, CSS and JavaScript — zero frameworks, zero templates — shipped through an agentic workflow: three competing design concepts prototyped live, headless-browser visual QA on every iteration, then this MANIFEST × LEDGER blend chosen and cut in a single working session.

AI appliedAgentic build loop — spec → concept prototypes → automated screenshot verification → ship · Claude as design & engineering partner
Under the hoodVanilla JS · IntersectionObserver reveals · CSS-only marquees · reduced-motion support · no dependencies
StatusLive — and iterating
Agentic WorkflowRapid PrototypingAutomated Visual QAVanilla JSMotion Design

Shipping Log

FOLIO 02

Public repositories, pulled straight off the docks at github.com/RafaelPupio and gitlab.com/RafaelPupio. Repo descriptions stay in English.

Last sync: 2026-10-03 · Auto-refreshes daily
Self-Updating

rafaelpupio.com

HTML / CSS / JS · GitHub

This site. Hand-written HTML/CSS/JS with zero dependencies, seven languages switched at runtime, and a GitHub Actions pipeline that pulls the GitHub & GitLab APIs daily, HTML-escapes everything it fetches, auto-commits, and redeploys through Vercel.

GitHub ActionsVanilla JSi18nCI/CDVercel
Inspect →
★ 1

ChurchChatBox

TypeScript · ★ 1 · GitHub

Secretaria Virtual — a WhatsApp church secretary bot (pt-BR). Automates the secretarial front desk over the channel Brazilian communities actually use. Next.js · PostgreSQL · Drizzle ORM.

WhatsApp BotNext.jsPostgreSQLDrizzlept-BR
Inspect →
New Cargo

Agents

Markdown · GitHub

Reusable AI agents as self-contained, portable Markdown modules — no framework, no lock-in. First aboard: ecoprompt, an agent that turns rough ideas into refined briefs and optimizes costly prompts for coding agents. Each agent ships with its own system prompt, knowledge files and skills, portable across Claude Code, Claude Projects and Custom GPTs.

AI AgentsPrompt EngineeringClaude CodePortable SkillsMIT
Inspect →
On Watch

service-watchdog

Python 100% · GitHub

Single-file, cross-platform service monitoring. Checks what you tell it to check, repairs only what is actually down, and reports outward — a dead-man's switch means a dead machine still raises an alarm.

DevOpsSREMonitoringSelf-Healing
Inspect →
Shipped

purged-walkforward

Python 100% · GitHub

Purged, embargoed walk-forward cross-validation for time series with overlapping labels. Shuffled K-fold lies on temporal data — look-ahead bias and label-overlap leakage inflate scores. This is the NumPy-only, scikit-learn-compatible fix: walk-forward ordering, purging, embargo periods, uniqueness weighting. Honest validation for models that trade.

Machine LearningTime SeriesCross-Validationscikit-learnNumPy
Inspect →
Live

landing-page-cav

TypeScript · GitHub

Landing page for Comunidade Árvore da Vida (Lucas do Rio Verde, MT). Next.js 16 · React 19 · TypeScript, with all content in a single Zod-validated JSON config, a dev-only content editor, SEO and full accessibility. Tested and deploy-ready.

Next.js 16React 19ZodSEOA11y
Inspect →
Opening

GitLab · @RafaelPupio

New berth · GitLab

Manifest just opened at this port. Projects will appear in the log as they ship — this page re-checks the docks every day.

GitLabIncoming
Inspect →
◤ Classified CargoPrivate holds · tools & process declared · contents sealed · inspection via contact
Classified

Vessel 01 · KX-94 DELTA

ML Trading Programme · Python · MQL5 · 3 private repos

An end-to-end algorithmic trading programme built on one thesis: history's famous algo-trading disasters were operational failures at machine speed, not bad forecasts — so the model is never allowed to be the last line of defense.

Signal layerPython · pandas / NumPy feature pipelines · gradient-boosted tree models · news & sentiment ingestion · scheduled retraining behind an out-of-sample validation gate · ONNX-ready deployment
GovernanceMQL5 Expert Advisor on MetaTrader 5 that treats every signal as untrusted: risk-% position sizing, hard daily-loss cutoff, signal-staleness guard, news-blackout windows, human kill-switch, append-only audit trail
SciencePre-registered cross-market replication (forex → B3 index/dollar futures) · frozen rulesets · live trials register · statistical power analysis before any edge claim — validation tooling open-sourced as purged-walkforward
ArchitectureModel and execution run as two isolated processes over fault-tolerant IPC — no forecast can bypass a risk limit
PythonMQL5 / MT5Gradient BoostingFeature EngineeringRisk GovernanceExperimental Design
Classified

Vessel 02 · MR-7 SIGMA

AI Finance Platform · TypeScript · Python · Monorepo

A personal + business finance platform whose core engineering problem is trust: LLMs are probabilistic, money is not.

InvariantLLM document parser (PDF / OFX bank statements) wrapped in an arithmetic proof — an import is rejected unless the extracted entries reconcile to the statement's own declared closing balance, to the cent
StackTypeScript monorepo · Expo / React Native client (Android + iOS) that renders but never computes · pure-TS financial core with signed integer-cent arithmetic (zero floats) · Python FastAPI ingestion service · PostgreSQL on Supabase
SecurityDefault-deny row-level security on every table with user-scoped policies · dedicated RLS test suite against Postgres in Docker · anonymized fixtures only — no real financial documents ever enter the repo
ProcessSpec-first development — approved design specs and implementation plans versioned in-repo · GitHub Actions CI
LLM ParsingFastAPIReact Native / ExpoPostgreSQL · RLSSupabaseTypeScript
Classified

Vessel 03 · BQ-40 NOVA

Social Habit Platform · Web + Mobile

A group-challenge product shipped twice — web and native — with the engagement mechanics designed from scratch, not copied.

WebVite + React SPA · component-state architecture · zero-backend demo mode seeded through localStorage
MobileExpo / TypeScript React Native sibling app sharing the product logic
MechanicsStreak algorithms · tie-aware leaderboards · daily-rotating content · reaction feeds · group statistics
OpsGitHub Actions CI on both repos
ReactViteExpoTypeScriptProduct EngineeringCI
Classified

Vessel 04 · CB-1 ORACLE

Agentic Infrastructure · Cross-machine

The infrastructure that makes every project above move faster. AI coding agents are stateless — so their memory has to be engineered.

MemoryVersion-controlled Obsidian vault as durable agent context — decision logs, status files, and research notes, all cross-linked
HygieneLog-rotation policies tuned to token budgets · present-tense-only status files · single-source agent configuration shared across Claude Code and other coding agents
BootstrapOne shell script restores the complete agent environment on any machine · a handoff protocol lets any fresh session resume mid-project with full context
Context EngineeringAgent MemoryObsidianShellGitClaude Code

Method & Stack

FOLIO 03

The tools, layer by layer — every line backed by a public repo you can open.

AI / MLDeep-learning agent with a daily retraining loop (trading pipeline) · RAG over PostgreSQL + pgvector embeddings with source citations · multi-stage document ingest with extractor → verifier agent chains · tool-using agents on the Vercel AI SDK (knowledge search, calendar, human escalation) → MORDOMO, trading agent
ML RigorPurged & embargoed walk-forward cross-validation for overlapping-label time series — look-ahead bias and label-leakage control, uniqueness weighting; NumPy-only, scikit-learn-compatible API → purged-walkforward
LanguagesPython for ML and systems tooling · TypeScript end-to-end for product code
Backend / DataPostgreSQL (+ pgvector) · Drizzle ORM · Next.js server routes · Zod-validated typed configs → MORDOMO, ChurchChatBox, landing-page-cav
FrontendNext.js 16 · React 19 · SEO & WCAG accessibility · dependency-free motion design in vanilla JS/CSS → landing-page-cav, this site
Agentic Eng.Claude Code as engineering partner — spec → persistent .md context files → plan-first build → headless-browser visual QA on every iteration → every repo above
Ops / SRECross-platform service monitoring with self-healing restarts and a dead-man's switch — a dead machine still raises an alarm → service-watchdog
rafael@pipeline: ~ $ cat METHOD.md
01 SPEC     — define the mission like an operator: scope, constraints, definition of done
02 CONTEXT  — engineer persistent .md knowledge files so every session starts fully briefed
03 BUILD    — pair with Claude Code: plan first, generate, review every change
04 VERIFY   — tests, screenshots, metrics — ship only what survives inspection
▸ edge: 10+ years inside administration, operations & financial operations.
▸ I don't build AI for demos — I build it for domains I've actually run.
$

Credentials

FOLIO 04

What each course covered, and where the tools are applied. Full record on LinkedIn — /in/rafaelpupiovieira. Credential details stay in English.

DiplomasFormal education
4.1
Master of Business Administration (MBA)

Pontifícia Universidade Católica do Rio Grande do Sul, 2024–2026. Strategy, corporate finance, operations and people management. Applied to how projects are scoped and prioritised: unit economics, business cases, and stakeholder-facing documentation for every build.

Verified
4.2
Associate Degree in Full-Stack Software Development

2020–2022. Two-year degree in front-end, back-end and database engineering. Applied across the public repos: TypeScript, Next.js and React, PostgreSQL with Drizzle ORM, REST and server routes, Zod-validated configs, and CI on GitHub Actions.

Verified
4.3
Bachelor's Degree in Theology

Fatin, 2012–2016. Exegesis, hermeneutics, historical research and academic writing. Applied as domain expertise behind MORDOMO and ChurchChatBox — knowledge-base structuring, source-cited RAG answers, and pt-BR content design for church operations.

Verified
Certifications & LicensesNewest first
4.4
Generative and Agentic AI

Saïd Business School, University of Oxford, 2026. Generative models, agent design patterns, orchestration and evaluation. Applied in MORDOMO and my agent infrastructure: tool-using agents on the Vercel AI SDK, extractor → verifier chains, and version-controlled context files that keep stateless models briefed.

Verified
4.5
AI Fundamentals in Financial Services

Saïd Business School, University of Oxford, 2026. Machine learning and AI applied to financial services: model capabilities and limits, data requirements, risk and governance. Applied in my trading programme — pandas / NumPy feature pipelines, gradient-boosted models, and out-of-sample validation gates before any model reaches execution.

Verified
4.6
Marketing Intelligence, Data Analytics and Agile Management

2026. Market and customer intelligence, data analytics for decision-making, and agile delivery management. Applied to how the work is run: metrics-first scoping, iterative delivery reviewed at every step, and the analytics behind the automated weekly reporting in MORDOMO.

Verified
4.7
Claude 101 & Claude Code 101

Anthropic, 2026. Prompt design, tool use and agentic development with Claude Code. Applied daily: spec-first builds, persistent Markdown context files, subagent delegation, and headless-browser visual QA on every iteration.

Verified
4.8
Private Pilot License

Aeronorth Curso de Piloto Privado, 2025 — 50+ hours of real flight. Meteorology, navigation, air law, aircraft systems and checklist-driven procedure. Applied directly to production discipline: pre-flight-style checklists, hard operating limits and explicit abort criteria — the model behind the kill-switch and daily-loss cutoff in my trading system.

Verified
4.9
Foundations of Cybersecurity

Google, 2025. Security domains, the CIA triad, common threat types and the NIST frameworks. Applied to this site: every value fetched from external APIs is HTML-escaped before it reaches the page, no untrusted data touches innerHTML, and outbound links carry rel="noopener noreferrer".

Verified
Field RecordNot a credential
4.10
A Decade of Operations

2016–2026. FP&A, international logistics and business administration. Applied as the targeting layer: forecast and cash-flow modelling, process mapping, and the domain judgment that decides which workflows are worth automating.

Field-Tested