$ whoami

Aldair Cortés

Data & ML Platform Engineer

I get data and models into production, and I make sure they stay there.

I build data platforms and ML systems that run in production, and I lead the teams that keep them healthy. Pipelines, cloud infrastructure, engineering standards, and the occasional mechanical keyboard.

Based in Mexico · works in English and Spanish

Dataplatforms and pipelines that survive Mondays
MLfrom notebook to production, with tests
OpsCI/CD, IaC and monitoring on GCP, Azure and Databricks
Peoplestandards, code reviews and mentoring

## work

Case studies

Real systems, told without the confidential parts. The logs are staying where they are.

An ML platform on Databricks

CEMEX Global Data Science · 2022 – today

Problem
Models were deployed by hand, about an hour each, and every model had its own serving endpoint.
What I did
Designed an ML lifecycle standard on Databricks: automated CI/CD, MLflow pipelines, self-updating deployments and monitoring for both infrastructure and model drift. Rolled it out across 4 production projects.
Results
  • 67% faster deployments: 60 minutes of manual work became a 20-minute automated pipeline.
  • 70% lower inference cost by routing one endpoint to the model each request needs.
  • Infrastructure and model-performance monitoring on all 4 projects.
  • Databricks
  • MLflow
  • PySpark
  • Asset Bundles
  • Lakeflow
  • CI/CD

A Customer Data Platform from zero

El Puerto de Liverpool · 2022

Problem
A retailer needed its first production Customer Data Platform, built greenfield.
What I did
Led the launch: 4 Java/Spring Boot microservices, a data access layer, and streaming and batch pipelines on GCP, provisioned with Terraform. Real-time ingestion ran on Pub/Sub, Cloud Functions matched data and Cloud Run executed the pipelines that deployed updates.
Results
  • 75% lower Cloud SQL cost (60 to 6 vCPUs) by tuning queries, stored procedures and Apache Beam code.
  • Observability for services, pipelines and databases that cut the time to detect performance problems.
  • Near real-time Looker dashboards for daily sales and nationwide ticket volume.
  • GCP
  • BigQuery
  • Dataflow
  • Pub/Sub
  • Cloud Run
  • Terraform
  • Java

From R to Python, in four countries

CEMEX · supply chain services

Problem
Five critical supply chain services written in R had to become maintainable production software.
What I did
Led the migration to Python and FastAPI, deployed and monitored on Azure across the US, Mexico, Colombia and the UK. Also owned the dispatch scheduling algorithm behind concrete delivery logistics.
Results
  • 50% better response time on the migrated services.
  • Dispatch scheduling answers in under 5 seconds, with fewer manual validations and real-time inventory visibility.
  • Migrated legacy Airflow DAGs from Cloud Composer 1 to 2.
  • Python
  • FastAPI
  • Azure
  • Airflow
  • Cloud Composer

Engineering standards, human and AI

Technical leadership · 30 engineers

Problem
Data science work needed the same engineering rigor as software, and AI assistants needed house rules too.
What I did
Led the Testing Learning Path, set MLOps standards and code templates, and defined how the team works with GitHub Copilot and Claude Code: better context engineering and lower AI credit use.
Results
  • Test coverage from 0% to 70%+ across 14 data science projects.
  • Two mentees hired as junior ML engineers.
  • Co-organizer of the Mesa de Arquitectura, an internal architecture committee with its own Python and design patterns curriculum.
  • Pytest
  • CI
  • Copilot
  • Claude Code
  • SOLID
  • DDD

## projects

Public projects

Teaching material, research, geodata and hardware I could not leave alone.

Practical CKAN manual

A four-part hands-on manual for publishing open data with CKAN: Docker Compose install, FileStore and DataStore, the API, and a full data-to-portal workflow. In Spanish.

  • CKAN
  • Docker
  • Open data

// docker compose up && hope

Data governance playbook

An integrated data policy (quality, classification, access) for a sample company, plus a manual that turns it into reproducible OpenMetadata configuration: tags, RBAC, quality tests and lineage. In Spanish.

  • Governance
  • OpenMetadata
  • RBAC

// policy-as-code, minus the drama

SQL guide with exercises

From basic queries on a customer-orders dataset to indexes, query plans, permissions and backups in PostgreSQL. In Spanish.

  • SQL
  • PostgreSQL

// SELECT * FROM patience;

KAN interpretability

My master’s research at IPN on Kolmogorov-Arnold Networks and the interpretability of neural networks. Thesis in progress; these are the related forks.

  • KAN
  • Interpretability
  • Research

// interpretable, unlike my commit messages

Drug-drug interaction prediction

A graph machine learning course paper that reproduces DPDDI, a deep predictor for drug-drug interactions. Written in LaTeX at CIC-IPN. In Spanish.

  • Graph ML
  • LaTeX

// peer review, but make it a notebook

Human-inspired optimization

A Python implementation of the Social Algorithm tested on benchmark functions, with convergence and diversity analysis over repeated runs.

  • Python
  • Optimization

// converges. eventually.

Random walk robots

Random walk simulation and analysis with PyGame for a time series course at CIC-IPN, built with a classmate.

  • Python
  • PyGame
  • Time series

// built with caffeine and hurry (the README says so)

Course notebooks

Pattern recognition and deep learning notebooks from my master’s courses: logistic regression, SVM, back-propagation and deep graph learning.

  • Notebooks
  • Deep learning

// loss going down, mostly

Point-in-polygon on a map of Mexico

Automates counting the points that fall inside polygons with GeoPandas and draws them on a Bokeh map. Built for broadcast coverage analysis.

  • GeoPandas
  • Bokeh
  • GIS

// finally, a use for all those shapefiles

CanSat telemetry

On-board telemetry with Arduino, Xbee and a GY-80 sensor, plus a Python ground station with OpenGL, pygame and Matplotlib.

  • Arduino
  • Xbee
  • Python

// my first deploy went to space(ish)

Abogabot: from requirement to UI

A template for taking a software requirement from understanding document to user stories, flow diagrams, UML and UI design, shown with a case management platform for a law firm. In Spanish.

  • Requirements
  • UML
  • UX

// objection: scope creep

React Tetris

Tetris in React, from 2019.

  • React
  • JavaScript

// game over

## keyboards

Mechanical keyboard nerd

Diagnosis: chronic. Prescription: one more keyboard.

When I am not moving data, I am probably moving switches. I translated the README, the bill of materials and the build guide of Lumberjack into Spanish, so more people can solder their own board.

Lumberjack is a through-hole ortholinear 60% PCB designed by Paul James. All through-hole parts, no SMD soldering, and it works in standard tray-mount 60% cases. The design is his; the Spanish version is mine.

  • Through-hole
  • Ortholinear 60%
  • KiCad
  • MX switches

// There is also a QMK firmware fork on my GitHub, because of course there is.

$ ls ~/keyboards

  • LumberjackOrtholinear 60%, through-hole, the one I translated.
  • HHKB Hybrid Type-STopre, quiet, 60 keys. The one that started the arguments.
  • Lily58 ProSplit, 58 keys. Technically under 60%, which only helps the cause.
  • DisciplineA 65% with through-hole parts. The name is ironic.
  • Rainy 75 ProA 75%. We do not talk about the Rainy.

// Official position: the best keyboard is a 60%. Arrow keys live on a layer, and a layer is a state of mind.

## lessons

Lessons learned

Two lessons I learned the hard way, so you can skip the hard part. Or at least nod knowingly.

Start simple. Stay native. Earn the framework.

More than once I reached for the big framework before the problem had earned it, and paid for it in setup, upgrades and onboarding. Complexity is a loan, and the interest shows up in every sprint. Over-engineering feels like diligence while you are doing it.

Rules I keep now

  • Start with what the platform already gives you: plain SQL, the standard library, native browser APIs, a script and a scheduler.
  • Prove the value with the simple version first. Add a framework only when a specific, measurable pain asks for it.
  • If a feature needs a diagram to justify its existence, ask whether it needs to exist.
$ ls node_modules
ls: cannot access 'node_modules': No such file or directory
# this site, on purpose

If it is not documented, it does not exist.

The most frequent victim of my undocumented decisions is future me, who opens a file and asks who wrote this. Knowledge that lives in one head is a single point of failure, and that head also takes vacations.

Rules I keep now

  • Write down the why, not only the how: the decision, the trade-offs and what was rejected.
  • Keep docs next to the code (README, decision records, runbooks) so they change in the same pull request.
  • A task is done when someone else can run it without asking me.
$ git blame
future-me: who wrote this?
git: you did.

## teaching

Teaching and community

I like it when other people’s pipelines stop breaking.

Private instructor

Databases, data governance and Python for government institutions. The CKAN, SQL and governance materials above are the open version of that work.

Mesa de Arquitectura

Co-organizer of an internal architecture committee. I wrote a Python onboarding curriculum (Clean Code, OOP, SOLID, design patterns) and a 16-topic roadmap covering DDD, hexagonal architecture, messaging, RAG, multi-agent orchestration and LLM evaluation.

Mentoring

Code reviews, pair programming and mentoring sessions for junior engineers and data scientists. Two of my mentees were hired as junior ML engineers.

## stack

Tools I reach for

Languages

  • Python
  • SQL
  • PySpark
  • Java
  • R
  • Shell

Data

  • Apache Beam
  • Airflow
  • Databricks
  • Spark
  • BigQuery
  • PostgreSQL / PostGIS
  • Snowflake

Cloud

  • GCP
  • Azure
  • Terraform
  • Docker
  • Kubernetes
  • CI/CD

ML & AI

  • MLflow
  • XGBoost
  • Prophet
  • Azure OpenAI
  • Gemini
  • MCP
  • GitHub Copilot
  • Claude Code

Backend

  • FastAPI
  • Spring Boot
  • Flask
  • Django
  • Looker

## contact

Let’s talk

Data platforms, MLOps, or mechanical keyboards. I answer all three.