How to Use Jev with Claude Code and Codex for Pix, SPEI, and Bre-B Operations

A LATAM fintech support queue doesn't speak one language. A Monday inbox mixes "mi SPEI no ha llegado", "o Pix caiu na conta errada", "¿cómo registro mi llave Bre-B?", a merchant writing in English about a settlement, and a customer who mentions PROCON or the CONDUSEF in the second paragraph. They all wait in arrival order, and the missed payroll transfer sits behind forty password resets.

This is a practical guide for payment operations and support leads at English-speaking LATAM fintech teams, and the engineer helping them: how to use Jev to triage Spanish and Portuguese tickets on Pix, SPEI, and Bre-B the moment they arrive, and how to start a Claude Code or Codex agent only on the tickets that need one. It builds on our general Jev triage guide, which covers the pattern itself. This one is about the rails, the languages, and the questions that change in the region.

What it does. Every ticket gets a category, a severity, and sensitive-case flags in under a second. Urgent payment problems start an investigation right away, routine questions are batched, and MED cases, chargebacks, and regulator complaints go straight to a person.

What it is not. Jev doesn't investigate, explain, or reply, and nobody has published how well it reads Spanish or Portuguese. It's a routing layer you validate on your own tickets before anything depends on it.

Jev in one paragraph

Jev is TypeSafe AI's first model, in early access since September 15, 2026. It reads the text you send and answers typed questions with probabilities: a choice from a list, a score across levels you describe, or a noul, the probability that a statement is true. TypeSafe lists 70 to 500 milliseconds end to end and $0.042 per million input tokens, with output free. Those are vendor numbers, so benchmark them on your own traffic. The general guide covers the question types and the Jev versus Claude Code comparison in detail.

The language question

This is what's different for LATAM teams, so start here.

TypeSafe's documentation doesn't list supported languages. When it released Jev 1.14 on October 14, 2026, it wrote that it expected the largest gain of that version in multilingual support, without publishing per-language numbers. So the honest position is: it may work well on your Spanish and Portuguese tickets, and you need to measure it before you rely on it.

Three things to test explicitly:

The queue on LATAM rails

Notice that the right answer for several rows is "no agent." Fraud cases on Pix are time-bound under the MED rules, and regulator complaints have their own clocks. Jev's job there is to make sure they never wait behind routine tickets.

How it works

1. Your team writes the questions and routing rules. Categories, severity levels, and what always goes to a person.
2. Jev scores each ticket on arrival. All questions in one request.
3. Code applies the rules. Start an agent now, batch it, or route it to a named person.
4. The agent investigates in Runtime. Read-only access, the runbook for that category, and a draft in the customer's language.
5. A person approves. Anything that moves money or reaches a customer waits for sign-off.

The stack

Runtime is the operating system for AI agents. Each agent runs on its own isolated computer, wiped at the end of the session, with its own scoped credentials, and waits for a named person's approval before moving money or replying to a customer. CPF, CURP, account numbers, and cards are masked before they reach the model or the logs, every run is recorded, and a Jev routing rule can start a session directly. It can run in your own AWS, GCP, or Azure account, which matters if your data needs to stay in Brazil or Mexico.

Step 1: Write the questions (human)

Jev reads your questions literally. TypeSafe's guidance is one specific thing per question, combined in code. For a LATAM payments queue, a first set:

Name the local institutions inside the question text. "A regulator, consumer authority, or lawyer is mentioned, such as CONDUSEF, the Superintendencia Financiera, Procon, or consumidor.gov.br" gives Jev the words your customers actually write.

Compute identifiers, amounts, and times in code. TypeSafe's documentation says Jev struggles with numbers and date comparisons. Extract the SPEI tracking key, the Pix EndToEndId, the amount, and how long ago the transfer happened in code, and pass the results as plain sentences in the state.

Step 2: Write the routing rules

Set the numbers from your history, per language. Keep the sensitive thresholds low: a fraud analyst reading a few false alarms costs minutes, a missed MED case costs far more.

Step 3: Build the triage with Claude Code or Codex

TypeSafe has Python and JavaScript SDKs and publishes an agent skill for Claude Code and Codex. Put your TypeSafe API key in your secrets and ask the agent to build it.

"Build a triage service. For each new Zendesk or WhatsApp ticket, detect the language, extract in code the SPEI tracking key, Pix EndToEndId, Bre-B reference, amount, and time since the transfer, then send the ticket text and those facts to Jev with the questions in triage-questions.md. Apply the rules in routing.md. Log every Jev answer with the ticket ID, language, model version, and rule that fired. Test on the 600 labeled tickets in /fixtures and report agreement separately for Spanish, Portuguese, English, and mixed-language tickets. Then run the same test with the Spanish and Portuguese versions of the questions and compare."

Two habits make these prompts work harder:

• Say why, not just what. "Report agreement per language, because a strong English score can hide a weak Portuguese one" gets you the breakdown you actually need.
• Name the failure you're preventing. "A ticket that describes a scam must never start an agent, even with low severity" exists because that's the MED case that misses its deadline.

Step 4: Set up the investigating agent

In Runtime, create the agent the triage service will start: read-only connections to your ticketing tool, core ledger, payment provider, and logs, and one runbook per Jev category. Add credentials through the dashboard, never by pasting them in chat. Then test the limits: ask the agent to refund a charge or send a Pix and confirm it can't.

Step 5: Let Jev start the agent

"When the routing rule is 'investigate now,' start a session with the ticket, the Jev answers, and the category. Load the runbook for that category. Find the transfer using the tracking key, EndToEndId, or reference, cite every record you used, and post to #payment-ops a draft reply in the customer's language and a recommended action. Never reply to the customer or move money."

Run it in shadow mode for one to two weeks: Jev routes, your team still triages, and you compare. Go live one queue and one language at a time.

What stays human

Lessons

• Measure per language from day one. Blended accuracy is the number most likely to mislead you.
• Fix the question before blaming the model. Jev reads literally, in any language.
• Ticket text is untrusted. TypeSafe notes Jev can be misled by adversarial content. No Jev score should grant an agent more access or skip an approval.
• Pin the model version. Each response includes the versioned model ID. When it changes, rerun your per-language test set before trusting your thresholds.

What compliance will ask

• Who owns the triage and the agent. A named owner per queue and scoped permissions per agent.
• That nothing moves on its own. Refunds and replies wait for a person.
• What data leaves. Treat TypeSafe like any third-party processor under the LGPD in Brazil or your local data protection law, and send only what each question needs. Runtime masks sensitive data before it reaches the model and never trains on customer data.
• Why a ticket was routed the way it was. Log every Jev answer next to the ticket, alongside Runtime's record of each investigation.

Measuring it

A realistic timeline

Rain, a card issuer present in more than 150 countries, uses Runtime for technical support, fraud and risk monitoring, compliance, cybersecurity, engineering investigations, PR reviews, QA testing, and reporting, and has saved $750k in vendor and salary spend.

The point

Cheap decisions let you score every ticket the moment it arrives, in Spanish, Portuguese, or English. Whether Jev reads your languages well enough is a question your own tickets can answer in a week. Once they do, Jev sorts, the agent investigates in Runtime, and your team makes the calls that need a name.

For the Spanish version of this pattern, see Jev para operaciones de pagos. For onboarding and AML queues, see Jev for KYC, KYB, and AML.

Back to the tickets.