I BUILD AI
SYSTEMS THAT
SURVIVE PRODUCTIONTGT LOCK.
I build AI automation and data pipelines that show their work and stay reliable in production. A multi-agent qualifier where two models cross-check each other and abstain when they disagree. An agent that writes SQL and corrects its own errors. A RAG pipeline that measures its retrieval quality before it answers. MCP servers that let an AI assistant query live business data mid-conversation.
Underneath all of it is the data craft that feeds every AI system: web scraping, PDF extraction, and Python pipelines that turn messy sources (websites, PDFs, public records, scattered APIs) into clean, structured data you can actually use.
All of it is public. Every claim on this page links back to its source: the repositories, a live app, published security advisories, a book you can download. Proof, not promise.
RETAIL MANAGEMENT SAAS
A full retail platform: fast point of sale, real-time inventory, integrated financials, a public online catalog, and advanced reports (margin, ABC curve, sales projection). Reads an entire Excel workbook on import. The interface ships in English and Portuguese. Built, shipped and operated end to end.
6 NA REDE
The official app of a footvolley club in Niterói: training schedule with attendance, balanced doubles draw, win rankings, events and gallery. A Next.js PWA shipped as a native iOS app via Capacitor, one codebase through Apple review, published under my own developer name.
TRUST-FIRST LOCAL SERVICES
A local-services marketplace for Brazil’s countryside towns: residents hire verified nearby workers for cleaning, small repairs and furniture assembly. Nobody appears in a search before passing full KYC: document capture, dual liveness selfies, automated face match against the document, and three criminal-record registries checked automatically, at under R$1 per verified worker. Requests typed or spoken in plain language are parsed by Claude. Automation flags, a human decides: the system never rejects anyone on its own. Privacy by design under LGPD: exact addresses and phone numbers never enter the system. Co-founded with the local founder who leads product; the entire technical build is mine.
CLIENT MISSIONS SHIP UNDER NDA. DESCRIBED BY OUTCOME, NOT BY NAME. QUOTES FROM PUBLIC REVIEWS OF COMPLETED CONTRACTS. ONE ENTRY IS A NAMED, PUBLIC COLLABORATION, NOT CLIENT WORK.
MARKET INTELLIGENCE PIPELINE
Daily scraping and enrichment for a US real-estate firm. Scheduler running 24/7 on a dedicated VPS, unattended.
His communication was excellent every step of the way.
CONFIG-DRIVEN CLEANSING ENGINE
Data cleansing driven entirely by configuration, backed by automated tests, so rules change without breaking pipelines.
Strong architecture, clear documentation, automated tests.
PROFESSIONAL REGISTRY EXTRACTION
Extraction and validation at scale: 7,010 validated records from a public physician registry, delivered clean and ahead of schedule.
Completed the work ahead of schedule and with accuracy. I would hire him again.
PRINT-VENDOR PRICING APIS
Six print vendors unified behind one pricing interface: a single contract over fragmented sources.
Clear, intelligent communication. Detail oriented.
MEDICAID FORM AUTOMATION
A manual government-form workflow automated end to end: fewer hands, fewer errors.
LEGAL RESEARCH DATA COLLECTION
Two years of public forum discussion, 3,221 threads and 58,686 comments, delivered with a completeness report the client reconciled against the raw file himself.
Vinicius worked with me on focusing the project scope before carrying out seamlessly.
REGULATORY REGISTER MONITORING
A regulator's public licence registers monitored end to end: extraction, change detection, QA, recurring controls, and an independent handover.
Built with handover and maintainability in mind, rather than creating something dependent on him.
MEP COST ESTIMATION ENGINE
A custom costing engine for MEP estimation: masterdata, templates and quantity takeoffs resolved deterministically, reproducing the client's reference workbook to the cent, 111 of 111 rates.
Understood the requirements quickly, delivered fast, and handled every piece of feedback seriously. I would gladly work with him again.
TRUSTWORTHY AGENTIC COMMERCE REVIEW
Invited technical reviewer on a decision-centered reference architecture for trustworthy agentic commerce by Dimitrios S. Sfyris, founder of AspectSoft, now published on arXiv. Credited by name for refining the separation between commercial eligibility and actor authority, the treatment of payment artifacts as evidence rather than automatic permission, and the controls that keep AI-generated claims grounded, scoped, and verifiable.
◂ READ THE PAPER ON arXiv →THE OTHER HALF OF BUILDING SYSTEMS THAT SURVIVE PRODUCTION: FINDING THE CRACKS IN EVERYONE ELSE'S. REAL, EXPLOITABLE, WITH A WORKING PROOF OF CONCEPT. IF THE IMPACT CANNOT BE PROVEN, IT DOES NOT BECOME A REPORT.
Every audit runs the same sequence. Nothing advances on a hunch. The AI-augmented pipeline widens the reach; the judgment of what is real stays with me.
DISCLOSURES
INCL. FRONTIER LABS
EVERY REPORT A REAL FINDING
SECURITY VULNERABILITY · ANTHROPIC
Found and reported a security vulnerability to Anthropic through their official bug bounty program on HackerOne. Reproduced with a working proof of concept, triaged, and awarded a bounty. Reported responsibly before the change reached a released tag.
AUTHORIZATION BYPASS · DIRECTUS
Row-level authorization bypass in Directus, CWE-863. The GitHub Security Advisory is still in coordinated disclosure, and I am credited on it as a collaborator. The mechanism stays private until it ships.
SSRF · MCP INSPECTOR
Server-side request forgery in the official Model Context Protocol Inspector, reported by me. The GitHub Security Advisory is still in coordinated disclosure, so the mechanism stays private until it ships. Directly in the AI-agent supply-chain surface.
PATH TRAVERSAL · MULTIPASS
Path traversal in Multipass, the virtual machine manager by Canonical, on the recursive file transfer path: a transfer could write outside its destination directory on the host. Reported by me, accepted by Canonical in September 2026, with the fix landing in the next feature release. The GitHub Security Advisory names me as reporter and goes public with that release.
SECURITY VULNERABILITY · MONGODB
Found and reported a security vulnerability in Evergreen, MongoDB’s open-source CI platform, through MongoDB’s bug bounty program on HackerOne: an authenticated user holding no roles could make the platform act with its own credentials against targets of the user’s choosing. Reproduced three times in an isolated rig, with no request reaching GitHub or MongoDB infrastructure and every artefact hashed. Validated, rewarded and resolved by MongoDB’s security team in September 2026.
OUT-OF-BOUNDS READ · CRUN
crun is the OCI runtime Podman and CRI-O run containers with. Its passwd parser tested for a digit with **s - '0' < 10, a signed comparison with no lower bound, so 186 of the 256 possible byte values kept the loop running and the null terminator was one of them. A passwd line whose uid field is empty walks the parser past the end of the buffer. The correct form came from musl in June 2020 and lost its lower bound 43 minutes later, in a commit that dropped one character to silence a compiler warning on the Alpine build, which is where the affected path compiles. No advisory, no CVE, no severity assigned. It was fixed in the open as a correctness bug.
| AR-01 | ANTHROPIC | Arbitrary file read via PR-controlled symlink · CWE-22 · bounty paid | MEDIUM · 5.9 |
| AR-02 | DIRECTUS | Row-level authorization bypass · CWE-863 · credited collaborator · advisory in coordination | MODERATE · 6.5 |
| AR-03 | MCP INSPECTOR | Post-auth SSRF · reported by me · advisory in coordination | MODERATE |
| AR-04 | MULTIPASS · CANONICAL | Path traversal in recursive file transfer · CWE-22 · accepted, credited reporter · advisory publishes with the fix | MEDIUM · 6.5 |
| AR-05 | MONGODB | Credentialed write aimed by an unprivileged user · CWE-610 · bounty paid | LOW · RESOLVED |
| AR-06 | API GATEWAY · OSS | RCE via script sandbox escape · constructor bypass | CRITICAL |
| AR-07 | FRONTIER AI LAB | Arbitrary OS command exec via untrusted MCP server config | HIGH |
| AR-08 | WEB FRAMEWORK · OSS | Middleware auth bypass · incomplete-fix of a prior CVE | HIGH |
| AR-09 | API GATEWAY · OSS | RCE · trust-grant bypass of the STDIO safety modal | HIGH |
| AR-10 | ID-VERIFICATION VENDOR | Unauthenticated cross-tenant disclosure of ID-verification sessions | HIGH |
| AR-11 | AUDIO STREAMING PLATFORM | Cross-entity authorization bypass via path traversal | MEDIUM |
| AR-12 | DATABASE PLATFORM · OSS | Arbitrary file write via path traversal · zip-slip | MEDIUM |
| AR-13 | ID-VERIFICATION VENDOR | Live analytics write-keys disclosed · arbitrary event injection | MEDIUM |
| AR-14 | DIGITAL BANK | Unrestricted API key + backend surface via public runtime config | MEDIUM |
| AR-15 | DIGITAL BANK | Public S3 bucket listing · 830 objects enumerable | MEDIUM |
| AR-16 | EMBEDDED TLS / CRYPTO STACK | Out-of-bounds read in ECC public-key parsing | MEDIUM |
| AR-17 | CONTAINER ORCHESTRATION CLIENT · OSS | TLS hostname validation bypass in custom-CA mode · bearer token sent to a mismatched server | HIGH · 7.4 |
| AR-18 | DATABASE KUBERNETES OPERATOR · OSS | Cross-namespace privilege escalation via unvalidated resource reference · operator provisions root on another tenant | CRITICAL · 9.6 |
ARTIFICIAL INTELLIGENCE IN PRACTICE
A free 84-page handbook on applied AI, written for developers who want to build with LLMs instead of just reading about them. Everything in it was tested by hand: from "what is a token" all the way to a working agent with tools, and what it costs to run.
A decade building data platforms inside one of Latin America's largest credit bureaus and two global consulting firms. On my own time, I ship products people pay for.
I work written-first: scope agreed in writing, decisions documented, delivery you can audit. The proof is on this page: running systems, tests that pass, contracts that closed at five stars.
One transmission is enough: what you need, where the data lives, what done looks like. I reply with questions or a plan, in writing.