About
I build data systems for a retail chain, and I am the person who operates them afterwards.
I'm a data and automation engineer at Mercaldas, a retail grocery chain in Colombia, where I build the systems behind forecasting, pricing and internal operations.
I built the weekly demand forecast for thousands of product and store combinations, a market price platform the commercial team takes into supplier negotiations, and an assistant that answers business questions with verified numbers instead of invented ones.
Outside work I build my own infrastructure, and I'm interested in how these same methods apply to energy data: load forecasting, distributed generation, and the reliability problems that come with both.
How I work
One year of professional experience, so I do not sell myself on years. These four habits are what the five case studies have in common, and the part I want to be pressed on.
Controls live in code, not in habits
A rule you have to remember is not a rule. When I need a guarantee I make it structural: one entry point with a test that fails if a second appears, a required CI gate instead of a reminder.
Green checkmarks are not evidence
The failures that last longest are the ones the reporting does not cover. The useful question is not whether the job succeeded but whether the output is still good, measured independently. That is what caught a silent production failure.
Every dependency pays its own weight
I run Dagster where orchestration earns it, and a hand-written scheduler where it does not. I built a graph retrieval layer, measured it against plain grep, and deleted it when it did not win. Sophistication that does not measure better is cost.
Write down the trade-off
Test coverage tracked as debt is a decision. The same gap undocumented is an oversight. Every case study here names what was given up and why.
What transfers to energy
The domain changes. Most of the engineering does not. Rather than ask a reader in that sector to do the translation, here it is.
| What I built in retail | What it is called in energy |
|---|---|
| Weekly forecast of thousands of product×store series with LightGBM and Dagster | Multi-series load and demand forecasting at scale |
| Ingesting a state statistics office's weekly bulletins with no API and a drifting format | Ingesting regulated market data: system operators, grid operators, meters |
| Medallion architecture on Delta Lake with per-layer assertions | A data platform with quality layers and traceable lineage |
| Expected versus actual price, with an automated written reading of each series | Deviation of actual versus expected generation, with automated reporting |
| Human-in-the-loop gates and adversarial review of my own AI guardrails | Reliability judgment for systems with a model in the loop |
| An audit that caught a silent production failure the job status could not show | Operational monitoring that measures the outcome, not the pipeline's own opinion |
A forecast that has to be right every week teaches the same lessons whether the series is demand or load: inputs go stale, monitoring lies to you, and the interval matters more than the point estimate.
Contact
Manizales, Colombia