I build the systems
data moves through.

Problem

Colombia publishes all public contracting through SECOP II, but the raw feed is unusable for analysis: values go unreported, entity names are inconsistent, and the volume makes ad-hoc querying expensive.

Decisions

• Modelled the warehouse as ELT rather than ETL — land raw in GCS, transform in BigQuery with dbt, so a bad transform never costs a re-extract.
• Pre-aggregated into marts, keeping only supplier concentration as a live fact-table query (~163 MiB per entity); the entity picker is served from memory instead.
• Treated missing data as a first-class value: a null renders as a hatched gap, never a zero, because a confident zero is a wrong answer where no answer exists.
• Encoded that distinction as texture rather than colour, so it survives greyscale, printing and colour vision deficiency.
• Row counts are read live from the API's /health endpoint, never hardcoded — the marts rebuild daily and the figures drift.

Problem

Breast histology classification is usually framed as one binary question, but clinicians need the subtype too — and a model that cannot show why it decided is not usable in a clinical conversation.

Decisions

• Chose a single multi-task head over two separate models: benign/malignant and the 8 clinical subtypes share the same features, so one backbone learns both and cannot contradict itself.
• Picked EfficientNet-B0 for its accuracy-per-parameter — the model had to fit a free-tier inference container, not a training cluster.
• Added Grad-CAM heatmaps so every prediction ships with the region that drove it.
• Paired the classifier with an AlphaFold2 structural comparison of PALB2 (wild-type vs mutant) in an interactive 3D viewer.