I'm Matías Roubaud, an engineer with 7+ years taking data to production across the full stack — from scalable AWS pipelines to executive dashboards and ML models. I currently lead a Data & AI function, and I help companies build the infrastructure they need to decide with reliable data.
I work end to end: from data architecture to the business decision. Four complementary areas so your data, your cloud and your IT work for the business.
Scalable, governed data pipelines that feed analytics and models with reliable data.
Executive dashboards that turn raw data into decisions for leadership.
ML models from idea to production, integrated into your decision processes.
Cloud infrastructure and IT/security operations across the Microsoft and AWS ecosystems.
Four pieces showing how I work end to end: from infrastructure to dashboard, and from the model to the metric that says whether it's any good. All built on synthetic data — no client information is used.
View full image ↗ Four upstream sources, a layered data lake (bronze / silver / gold) on S3, a Glue catalog and consumption through Athena, Power BI and Grafana — all provisioned with Terraform. One IAM role per source scoped to its own prefix, and freshness alarms on top of error alarms: a Lambda that stops running doesn't fail, it goes quiet, and the data ages until someone notices the dashboard is stale.
View full image ↗ Executive reporting for a DTC e-commerce business: 75,053 orders across two years, with revenue and margin broken down by month, channel, category and region. The value isn't the chart, it's what comes out of it: footwear returns run at double every other category, and cutting that rate by two points is worth about €32k a year in recovered margin.
View full image ↗ Gradient boosting on calendar and lag features, validated with rolling-origin backtesting across 6 folds against a seasonal naive baseline. The model isn't the point, the evaluation is: every lag is at least 28 days old, because on the day the forecast actually runs the last 27 days don't exist yet. The weakest fold beats the baseline by just 4.8% against ~30% for the rest: it predicts the weeks after the January sale having seen a single previous January. Error isn't uniform, and a mean MAPE hides where the forecast stops being reliable.
View full image ↗ Measures the part of a RAG system that decides whether it works: if the right passage never reaches the context window, no prompting recovers it. A 236-document knowledge base, 14 gold questions and four chunking strategies against the same retriever, so any difference is attributable to chunking alone. A third of the questions fail at retrieval, and every failure is a paraphrase — which points at query rewriting or a reranker, not at smaller chunks.
7+ years across analytics, data engineering, development and IT/security — in industry, sports and tech.
I lead the Data & Analytics team and define the analytics and AI strategy on AWS: governed pipelines, executive dashboards (Power BI / Grafana) and models on SageMaker.
Led 80+ installations and data projects across multiple regions, building analytics solutions in SQL, Python, Tableau and Power BI.
Sports performance metrics and dashboards using Python and BI tools.
Web applications with React, Node.js, AWS and GCP for various clients.
Industrial clients and automation solutions; market analysis and technical training.
Teaching and research in engineering.
I come from electrical engineering and that shapes how I work: I understand the business problem before writing a line of code. I've spent 7+ years turning data into value across the full stack — from AWS pipelines to executive BI and ML models in production. Today I lead a Data & AI function at Basquevolt: I set the strategy, build the cloud infrastructure and ship dashboards and models that drive decisions.
My path went through analytics, engineering and IT/security, which lets me see the full picture. I also founded FingTwice, an online academy for engineering students in Uruguay. Teaching forces me to explain the complex in simple terms — and that shows in how I deliver clear, well-documented solutions.
Production-proven technologies. I pick the tool for the problem, not the other way around.
Tell me what you need and we'll set up a first no-commitment chat. I reply within 24 business hours.
Thanks! Your message was sent. I'll get back to you shortly.
or write me directly