§ 00 · profilecuritiba · remote · pt / en

Natan Cardeal.

Implementation SpecialistFinancial ReconciliationAI applied to regulated operations

nowImplementation SpecialistSimetrikJun 2026 → present

beforeData SpecialistJones Software · NYCJul 2023 → May 2026

thesis

In regulated operations, automation is only worth what it can prove, and a discrepancy is something to report, never to resolve in silence.

on this page

01about
02study · AI-agent method
03skills
04path · 7 roles
06projects · indícia
07contact

/about · summary

summary

Regulated financial operations, with AI under governance.

I started in 2017 behind the counter of an insurance brokerage in Santa Catarina, opening claims for dented cars. Since then the problem that occupies me has stayed the same: two records that should say the same thing have to be proven equal, or have the difference explained. At the brokerage, it was the claim against the policy. At Jones, in New York, the Certificate of Insurance against the client's contractual requirement. Today, at Simetrik, it is the ERP receivable against what the acquirer scheduled, the settlement against the bank statement, the accounting balance against the ledger.

That is nearly ten years in regulated financial operations: insurance, insurance and real estate compliance and, since June 2026, financial and accounting reconciliation for Latin American clients. The tooling changed along the way, from spreadsheets to Python, from Python to LLMs in production from 2023, from LLMs to agents. The standard did not.

That standard became a method. At Simetrik, I execute the reconciliation configuration alongside an implementation agent that works under explicit rules: versioned instructions, human approval before any destructive action, the client's answer key as the acceptance test, source and date on every claim.

01·

~10

years in regulated financial operations

02highlight

2023

LLMs in production since

03·

~17k

insurance certificates audited, above 99.9% accuracy

04·

roles · 5 companies

/study · simetrik · method

2026 → presentfinancial reconciliation

case study

method · conceptual

Working with an AI agent under audit.

Within a few weeks, a reconciliation project piles up dozens of meetings, client documents, decisions that change course and platform measurements that age fast. At that volume the real risk is not missing information: it is an old assumption or an unchecked deduction circulating as fact and turning into configuration. The method keeps proof apart from interpretation. It describes process, with no client data.

flowhumanagent

01human
Instruction
scope and acceptance criteria written and versioned in the case's own repository
outputversioned instruction
02agent
Execution
the agent runs through the platform's command line, only on the declared source
outputraw output
03agent
Evidence
every claim states where it comes from: measured, documented, said, inferred or unknown
outputclaim + source + date
04human
Audit
I check against the client's answer key before the next step
outputapproved · rejected

rejected → back to instruction, with the reason logged

principles · where it applies

01Instruction02Execution03Evidence04Audit

01Proof ≠ interpretationwhat the client delivered, what was said and what the platform showed are recorded once; summaries derive from that

02Source and datea deduction never presents itself with the confidence of a measurement

03Logged decisionswhen one decision replaces another, the change is written down with the reason

04Human approvalno destructive action moves forward without approval and a way back

05Answer key as testthe client's own reference output is the acceptance test

Client indicators, volumes and configurations stay out under confidentiality. Method details in conversation.

see the role in the path

/skills · 4 groups

skills

Skills.

filter

01Reconciliation & Paymentsnew

Financial and accounting reconciliationSimetrik platform configurationAcquiring and sub-acquiringIssuing and closed-loop schemesPix and bank slipsAcquirer EDI filesBank statements and settlement filesReconciliation against the SAP ledgerStabilization validated with the client

02Applied AI & Data

Claude, Gemini and GPT in productionAgent governanceMulti-agent orchestrationOCR + LLM visionSchema-enforced structured outputLiteral evidence guardRAG over PostgreSQL, Qdrant and Neo4jPythonSQL (PostgreSQL)PlaywrightREST integrationsPower BISmartsheet API

03Insurance & Compliance

Insurance and real estate complianceCertificate of Insurance (COI) auditClient-driven regulatory frameworkClaims managementPolicy issuance and controlRisk analysis in technical linesBroker ↔ carrier intermediation

04Method & Tools

Business AnalysisTechnical-functional documentationProcess mappingScrumKanbanJiraClickUpGoogle WorkspaceQuiverGit and GitHub

/experience · timeline

7 roles5 companies

experience

Path.

From claims analyst at a regional brokerage to financial reconciliation at Simetrik.

20172018201920202021202220232024202520262027

SimetrikJonesDimensaE360Slomp

click a bar to open the role

§ 01 / 07role

Jun 2026 → presentRemote · Latin America · via PI3.AI

Simetrik

Implementation Specialist

Executing the configuration of financial and accounting reconciliations for Latin American clients, alongside an AI agent under human approval.

Financial and accounting reconciliationPaymentsImplementationGoverned AI agent

01I execute the reconciliation configuration on the Simetrik platform: taking the project from paper into the platform and validating, together with the client, that the result matches how the operation actually runs.
02At Natura, receivables from the consultant sales network against what the financial partner scheduled, settled and deposited. At Banese Card / Mulvi, one company acting as issuer, acquirer and sub-acquirer at once. At Alelo, accounting balances against the general ledger.
03When something does not match, the discrepancy is investigated down to its cause and taken to the client, never adjusted in silence to make it close.
04I work alongside an implementation agent under a method of my own: versioned context with source and date on every claim, recorded decisions, human approval before any destructive action, and the client's own reference output as the acceptance test.

/education · certs · languages

PUCPRSUSEPPT · EN

education