As the first ML Engineer in the region, I quickly adapted to the team and became one of the trusted voices in technical decision-making.

Bridging the Gap

Joining as the first engineer for LATAM, my initial goal was simple:
prove the level of software quality and delivery we could achieve from the region… and we did.

First Task: Deploying a Ray Serving PoC

My first assignment was to deploy a PoC using Ray Serve.
Not a bad starting point, especially considering that the long-term goal was to standardize all deployments on Kubernetes.

Luckily, we had a dedicated team managing our EKS clusters, so I focused on designing the ideal workflow:

Kuberay for ML inference & Processing

With this setup, we were able to support both bulk distributed inference for heavy processing and real-time inference for customer-facing services.

During the second iteration, we added Airflow integration for bulk processing, enabling idempotent pipelines with clearer orchestration.

Most challenging parts:

  • Cold starts of new nodes during autoscaling
  • Managing and observing failures when a worker node died suddenly

Second Task: Deploying NLP Classification & NER Services

We needed to ship two ML/NLP services:

  • Document Classification
  • Named Entity Recognition (NER)

Before touching any infrastructure, I evaluated cost, traffic patterns, and deployment options using the platforms available to us:

  • Amazon SageMaker
  • AWS Lambda
  • ECS Fargate
  • KubeRay (Kubernetes)
  • Triton Inference Server

As usual, data scientists prefer not to think much about infra, and the moment “NLP” is mentioned, everyone assumes “this needs GPUs”.
Spoiler: it didn’t.

My decision matrix

  1. Expected traffic patterns
  2. Whether traffic was continuous 24/7
  3. CPU vs GPU requirements and whether the models were even optimized for GPU
  4. Artifact size and model type

Final outcome

Everyone initially wanted KubeRay…
but we ended up deploying everything on AWS Lambda.

The models were lightweight some literally installable via pip so we shipped fast, cheap, fully serverless inference.

But it didn’t end there.

Because these services were customer-facing, we redesigned our Lambda deployment approach to make it:

  • fully standardized
  • reproducible
  • aligned with SRE requirements

→ “Lambda Deployment Best Practices” (internal doc)

Third Task: Improving Hiring for Mid/Senior Engineers

Part of my role involved rethinking how we evaluated Data Scientists, ML Engineers, and Data Engineers.

I iterated on a new framework focused on:

  • real-world engineering maturity
  • problem-solving in modern ML systems
  • MLOps fundamentals
  • RAG and distributed systems

Not leetcode tricks… actual engineering.

→ “How to Hire Senior Engineers” (playbook)

Fourth Task: Hackathons & Internal Tools

AI Medusa

A chatbot designed to guide DS/engineers on how to deploy