Marcello Oliveira

Now

I lead the AI product engineering team at Deel. This is how we build with language models today.

  1. 01

    LLM features in production

    We ship features that validate and generate job scopes and map job descriptions, inside products that people use to hire and pay teams around the world.

  2. 02

    A rules engine as the safety net

    The model proposes; a deterministic rules engine decides what can go through. When the model is wrong, the rules catch it before a person ever sees it.

  3. 03

    Structured output, validated

    Every model answer comes back as structured output and is validated against a schema before any code trusts it. Free text never flows straight into a product.

  4. 04

    One internal AI service

    We are moving model calls to Deel’s internal AI service, with batching and rate limiting built in, so every feature gets the same limits, logging and controls.

  5. 05

    Observable and tested

    Datadog shows latency, failures and usage per feature; Sentry catches the errors; Playwright end-to-end tests cover the AI flows like any other critical flow. Feature flags let us roll out and roll back safely.

  6. 06

    An AI-native team

    I lead the team to work AI-native: AI is part of the daily toolchain, the eval and the failure state are designed before the happy path, and prompts are reviewed like code.