Alfredo Estevez — Data Scientist, ML and AI Engineer

About

Alfredo Estevez sitting on a ledge overlooking the Roman Forum

Hey 👋🏼 I'm Alfredo, an AI Engineer based in Mexico 🇲🇽. I began with robots and circuits as a Mechatronics Engineer, then fell for machine learning during my Master's and never looked back. Today I spend my days building with foundation models and cloud tools. When I'm not coding, you'll probably find me exploring new places or trying the newest coffee shop ☕

Recent Projects

Global Content Translator

Enroute · Client: Integral Ad Science · 2026

Localizes IAS’s customer-facing Help Center from English into 12 languages and keeps every translation in sync whenever the English article changes.

Translation time cut from ~2 hours (human-involved) to ~8 minutes (p50) per article.
Glossary-constrained Gemini enforces ad-tech terminology and preserves rich-text structure (links, headings, tables). A bulk backfill was followed by a Cloud Run webhook that republishes all locales on every English edit.

Vertex AI
Gemini
Cloud Run
Firestore
Contentful
Python

Contentful in Gemini Enterprise

Enroute · Client: Integral Ad Science · 2025–2026

Lets employees ask Gemini Enterprise plain-language questions about Contentful content and get sourced, linked answers, without learning the Contentful API.

An MCP server on Cloud Run exposes CMS data as structured tools, and a Python adapter turns them into Google ADK functions for Vertex AI Agent Engine.
Implemented a read-only guardrail that prevents the agent from writing to Contentful, and every answer returns source URLs.

MCP
Google ADK
Vertex AI Agent Engine
Cloud Run
TypeScript
Python

Scientific Literature Assistant

Datyra · Client: El Corporatus · 2024

A scalable RAG agent that answers questions over scientific literature, combining document search with structured SQL data.

PDFs in AWS S3 are loaded, cleaned, chunked, enriched with metadata and embedded into Chroma for semantic search.
Agent orchestrated with LangGraph, including safety and fallback nodes.
Served through asynchronous FastAPI microservices in Docker on AWS Fargate.

RAG
LangGraph
LangChain
ChromaDB
FastAPI
AWS Fargate
Docker

Multi-Label Harmfulness Classifier

Datyra · Client: Fooji · 2023

A fine-tuned LLM that scores user-generated text across nine dimensions of online toxicity, so brand community managers can enforce brand safety.

94% F1 across all nine dimensions, with Llama-2 13B fine-tuned using QLoRA.
Trained on 500K texts from public datasets, APIs and custom scrapers, part of an 800K+ dataset spanning text, images and video.
Deployed on an Amazon SageMaker endpoint that Safesail calls for real-time toxicity scoring.

Llama-2 13B
QLoRA
HuggingFace
PyTorch
Amazon SageMaker

Experience

AI Engineer
Enroute
Client: Integral Ad Science · Monterrey, Mexico
Oct 2025 – Present
- Built an AI-assisted translation pipeline for a Contentful-backed product: a webhook-triggered service translates changed fields across 12 languages with Vertex AI (Gemini) and writes back via the Contentful Management API, cutting translation time from ~2 hours (human-involved) to ~8 minutes (p50) per article.
- Designed and shipped a multi-tenant AI assistant that answers natural-language questions about Contentful content, pairing an MCP HTTP server on Cloud Run with a Python agent on Vertex AI Agent Engine that routes per-user credentials through read-only operations.
- Python
- Vertex AI
- Gemini
- MCP
- Cloud Run
- Firestore
- Contentful
Data Scientist and AI Engineer
Datyra Inc.
San Diego, CA
Apr 2021 – Aug 2025
- Developed and managed a scalable RAG agent for scientific literature, combining structured SQL data, AWS S3 storage, asynchronous FastAPI services and Dockerized microservices on AWS, fed by a pipeline that extracts, cleans and vectorizes papers.
- Migrated relational data (PostgreSQL) to BigQuery and added materialized views, cutting query times by up to 80%, automating CSV report delivery and enabling internal reporting through Looker Explores.
- Built a data enrichment system on Google Cloud Functions that processed hundreds of Facebook leads, helping the marketing team allocate resources better.
- Led data collection and labeling strategy, building three databases totaling 800,000+ data points (500k texts, 200k images, 100k videos) to train a harm detection AI for Safesail.
- Fine-tuned a Llama-based LLM for harm detection, reaching a 94% F1 score across nine dimensions of text analysis, used by Fooji in its Safesail platform.
- Python
- FastAPI
- AWS
- Docker
- BigQuery
- Looker
- Llama
- LLM fine-tuning
Software Engineer
Visteon
Querétaro, Mexico
Aug 2019 – Feb 2021
- Developed semi-automated Python software to run functional and regression tests on Ford vehicle cluster software over CAN communication, reducing testing time by 60% and automating Excel test report generation.
- Python
- CAN
Junior Engineer
Technological University of Mixteca
Oaxaca, Mexico
Oct 2018 – May 2019
- Designed and simulated a chocolate tempering prototype machine using SolidWorks and MATLAB Simulink.
- Collected experimental data to build a regression model that accurately described the system dynamics.
- SolidWorks
- MATLAB
- Simulink
Junior Engineer
KATA, Solar Energy & Domotics
Mexico
Jul 2017 – Sep 2017
- Built a relational database of solar energy components to improve the accuracy and speed of budget generation.
- Developed an intelligent architecture on an STM32F4 microcontroller for a home automation gas monitoring system, enabling real-time detection and automated response to gas leaks.
- STM32F4
- Relational databases

Tech Stack

Programming

Python
C++
Go

AI & ML

TensorFlow
PyTorch
NumPy
Pandas
scikit-learn
Keras
HuggingFace
LangChain
LangGraph
LLMs
Prompting
MCP
Google ADK
RAG
MLOps

Cloud Platforms

AWS Lambda
AWS Fargate
AWS ECS
AWS CloudWatch
AWS S3
AWS SageMaker
GCP Vertex AI
GCP Firestore
GCP Cloud Run
GCP Pub/Sub
Azure
Docker

Databases

PostgreSQL
OrientDB
BigQuery
PgVector
ChromaDB
ElasticSearch

Tools

Jupyter
Linux
Looker
Tableau
REST APIs
FastAPI
Flask
Gunicorn
Contentful
Agile

Education

Degrees

Master of Science, Electronics Engineering and Artificial Intelligence
Technological University of Mixteca · 2019 – 2021
Bachelor’s in Mechatronics Engineering
Technological University of Mixteca · 2014 – 2019

Honors

CENEVAL National Award
National academic excellence recognition

Certifications

EF SET English Certificate
71/100 (C2 Proficient)
AI-Augmented Engineer, Professional
OOP and Algorithms with Python
ECMAScript 6+
Scrum Professional Course

Languages

Spanish
Native or bilingual
English
Full professional (EF SET C2)
Italian
Elementary

Master’s thesis

Development and implementation of a Mexican Sign Language sign identification system based on neural networks and the Jetson Nano embedded system

Universidad Tecnológica de la Mixteca, Huajuapan de León, Oaxaca · 2022

Open access at the UTM repository ↗

Publication

Implementation and comparison of the KNN and CNN methods for the recognition of static signs of the Mexican Sign Language

UNAM SOMI ICAT · Oct 15, 2021

Compares a classical method (K-Nearest Neighbors) with a deep learning method (Convolutional Neural Networks) for recognizing static signs of Mexican Sign Language.

View ISSN record ↗

Hi, I'm Alfredo Estevez