VINICIUS PEREIRA // FULL-STACK AI ENGINEER Verified record

Track record · Upwork & beyond

Systems that survive production. Delivered, reviewed, paid.

A perfect client rating across every engagement. A paid security bounty from Anthropic. Named credit in a published architecture reference. A fix merged into the container runtime behind Podman and CRI-O. Everything here is independently verifiable, none of it is a claim you take on faith.

0.0Average client rating
0+Open-source projects
0Security disclosures
$0KPaid by Anthropic

In writing · one contract, start to finish · 72 seconds

Who you're hiring

For more than ten years I've worked inside private companies on tightly-scoped problems, some involving government data, under confidentiality that outlasts the contract. That work stays private, and always will, so the engagements below are described by their shape, not their client names.

Discretion with sensitive data is part of what you hire, not a footnote to it.

01 Client reviews

Every engagement · 5.0 · verified on Upwork

★★★★★5.0 · Reference letter

Real-estate market intelligence

Public housing data turned into a ranked target list, scored and running live on a schedule. Delivered, then kept running on my own infrastructure.

BrandonVerified · Upwork
★★★★★5.0

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.

SteveVerified · Upwork
★★★★★5.0

Data cleansing at scale

A configuration-driven cleansing pipeline for a large product dataset. Same rules, any catalog, no hand-editing.

PaxleyVerified · Upwork
★★★★★5.0

Healthcare data extraction

7,010 physician records pulled clean from a bot-protected registry, structured and de-duplicated for direct use.

VanessaVerified · Upwork
★★★★★5.0

Pricing extraction

Live pricing pulled across multiple print vendors and normalized into one comparable structure.

EricVerified · Upwork
★★★★★5.0

Insurance form automation

A fillable Medicaid claim form (eMedNY), field-mapped and validated so it fills correctly every time.

JordanVerified · Upwork

02 Validation

External · verifiable · by name

Paid security bounty
$2,000

Anthropic paid for a vulnerability I found.

I disclosed a real security flaw through Anthropic's own program. They validated it and paid the bounty. It sits inside a broader track record of coordinated vulnerability disclosures across multiple programs, none dismissed.

Credited by name · arXiv
Trustworthy Agentic Commerce

Named reviewer on a published arXiv paper.

Invited as a technical reviewer on a decision-centered reference architecture for trustworthy agentic commerce, now published on arXiv. The paper credits me by name for sharpening how AI-generated claims carry their own limits: the propagation of inherited refusal through derived claims, and a cleaner separation of the axes each surface projects onto.

Read the paper on arXiv ↗
Merged upstream · containers/crun
6 years

A parser bug in the container runtime behind Podman and CRI-O.

crun's passwd parser tested for a digit with a signed comparison that had no lower bound, so 186 of the 256 possible byte values kept the loop running, the null terminator among them. A line with an empty uid field walked the parser past the end of the buffer. The correct version came from musl in June 2020 and lost that lower bound 43 minutes later, in a commit that removed one character to silence a warning on the Alpine build, which is where the affected path compiles. Traced back to that commit and fixed. Reviewed and merged by the project maintainer.

Read the merged pull request ↗

Receipts. Redacted where a coordinated disclosure is still in progress.

GitHub Security Advisory · MCP Inspector
GitHub security advisory on the MCP Inspector project, moderate severity, in triage, mechanism redacted

Reported by me, coordinated disclosure in progress. Mechanism redacted until it ships.

GitHub Security Advisory · Directus
GitHub security advisory on Directus showing a credited collaborator, mechanism redacted

Credited collaborator on the advisory. Mechanism redacted until it ships.

03 Selected work

Open source · clean-room builds · public on GitHub

/01 tenant-fence One corpus, no crossingRetrieval for a knowledge base shared by many customers, where the entitlement filter runs before the candidate set is scored rather than after it. The wrong version ships beside the right one so the suite can demonstrate the leak instead of describing it: same corpus, same question, same account, three sections returned one way and none at all the other. github ↗ /02 langgraph-production Agents measured, not trustedLangGraph agent with the layer that decides whether it can ship: routing scored against labelled fixtures, a human gate before irreversible actions, and a refund that survives a crash without being paid twice. github ↗ /03 bedrock Reliability under loadNatural-language-to-SQL data agent with a production stability harness. Built to close the works-in-test, breaks-in-prod gap. github ↗ /04 confidence-gate Ships only what it trustsValidates model output and scores confidence from external signals (grounding, agreement, schema), then routes each result to auto-accept, human review, or abstain. github ↗ /05 grounding-probe RAG that refuses to guessCounterfactual probe that proves a RAG answer came from its sources, not the model's memory. Catches the confident guess. github ↗ /06 doc-eval Documents into data, measuredField-level evaluation and a CI release gate for document extraction. A candidate that lifts the headline number while quietly regressing one field gets blocked. github ↗ /07 lead-quorum Multi-agent, in productionMulti-agent lead-qualification system deployed on Google Cloud Run. Awarded a Cloud Run community badge. github ↗ /08 scrape-sentinel Collection that notices changeTurns any scraper output into new, changed, and removed since the last run, with field-level deltas and an atomic snapshot. Silence stops meaning nothing happened. github ↗ /09 crosswatch Pipelines that run themselvesScheduled market-data pipeline: multi-source pull, scoring, and alerting on a cadence, with source drift accounted for. github ↗ /10 retell-sms-handler Voice that keeps its promisesIn-call SMS for a voice agent, fired deterministically from a conversation flow: consent gated, destination resolved in code, every send logged. github ↗

04 Beyond the platform

Products shipped & authored

Own product · production

DeskCenter

A production SaaS in daily use by more than 1,000 businesses. Built and maintained end to end.

Own product · shipped

Published iOS app

A live app on the Apple App Store, taken from web stack to store through a hybrid build.

Own product · pilot live

RendaPerto

A trust-first local-services marketplace for Brazil’s countryside towns. Full KYC before anyone is listed: document, liveness selfies, face match and three criminal-record registries, automated at under R$1 per verified worker.

See it ↗
Authored

AI in Practice

A free 84-page field manual on applied AI: RAG, agents, evaluation, and three end-to-end capstones.

Read it ↗