CentauroCentauro · AI engineering and security
Reliable AI agents.
Evidence for consequential decisions.
We build and assess AI agents and their execution layer—the harness that enforces limits, human approvals and inspectable evidence.
For teams answering to customers, auditors and risk committees.
Security · Reliability · Traceability
Centauro · Evidence in operation
Illustration · actions bounded by controls and human approval.
Trust is what makes
autonomy possible.
When an agent handles money, sensitive data or important decisions, trust becomes a condition for adoption and scale.
- CTO
Can we release it? What happens when it fails?
- CISO
What can it access or change? Can we investigate misuse?
- CFO / business
What exposure are we accepting? Who approves the action?
From an agent's action
to an inspectable decision.
See which information it used, which tools it called, which controls applied and where a person must decide.
Illustrative · synthetic dataExpected behavior only. No live system, executed tests or assurance conclusion.
Proposed action
Start with what the agent wants to do.
Invoice INV-104 · PO-88 · MXN 12,000. The agent retrieves the invoice and approved supplier record.
Proposed action
- Task
- Prepare supplier payment
- Source
- INV-104 / PO-88
- Amount
- MXN 12,000
Control decision
The boundary must hold, even when the agent asks.
The invoice names a different bank account. The execution service must hold payment, even if the agent requests it. The AP owner verifies the beneficiary. An authorized approver must approve the exact payment before execution.
Required decision
- Exception
- Beneficiary mismatch
- Action
- Hold payment for human review
- Approval
- Required for the exact payment
Recorded evidence
Leave a decision someone else can inspect.
Link source records, tool actions, policy version, timestamps, approval and final outcome in the case record.
Required evidence
- Trace
- Sources, tool actions, policy version
- Decision record
- Timestamps, approval and final outcome
Evaluate. Reinforce. Verify.
Start with a scoped assessment of one consequential workflow.
EvaluateDefine what good looks like.
Set the agent's task and acceptance criteria. Test its performance, permissions and response to errors and abuse.
ReinforcePut boundaries into operation.
Implement controls on tools, data and actions. Establish approvals, exception handling and accountable owners.
VerifyRetest. Keep the evidence current.
Document retest results and open risks. Agree review intervals and what must be checked when the system changes.
Know who stands
behind the work.
Miguel Kalifa · founder
Centauro is led by Miguel, with experience in internal AI management-system audit and readiness work. Scope, criteria and responsibilities are agreed at the outset.
Meet the founder →Bring the decision.
We’ll define the evidence.
Start with one use case and the decision you need to support.
A brief description is enough to start. Keep sensitive records out of the first email.