José Paz Rangel Rojas

A 300-property portfolio on one map. Surface area is measured geodesically from each parcel's own polygon instead of trusted from a field that was often missing or wrong, and acquisition prices are restated for inflation so a portfolio total means something. Built twice — R/Shiny and Python/Dash — over the same warehouse, plus a natural-language query box on a Snowflake Cortex agent.

What is this parcel worth, and how sure? A valuation model over 2,600 land transactions that returns a calibrated range, not a number — 90% promised, 90.2% measured. Prices are restated for inflation first, because otherwise the model learns that recent is expensive and calls it value. Pointed at the group's own portfolio, it finds the purchases made outside the market.

Leasing system for three shopping centres: floor plan, availability, tenants and history. Handing a unit back never deletes anything — it posts a negative-square-metre entry, so occupancy is the running sum and the history cannot drift from the current state. Drag the date slider and the plan redraws for any month since 2022, with no historical tables.

Nobody ever labelled a transaction as laundering, so a classifier is off the table. This learns what is normal for each client and ranks what departs from it. contamination comes from the team's actual review capacity, validation is an expanding window, and the evaluation is reported twice — with labels, and the way production will have to, without them.

24 entities pulled from a paginated ERP API into Snowflake from one declarative catalog, merged by row hash so a re-stamped modification date writes nothing. Includes a fingerprint reconciler, data-quality views and drift probes. The demo runs four passes against SQLite: pass three proves the merge is idempotent.

End-to-end reporting for Mexico's LFPIORPI: ERP to Snowflake to a Streamlit app behind AWS Cognito. The hard part was not the API — it was modelling partial payments, VAT and multi-invoice collections so the reported amount is the one that actually changed hands, instead of over-reporting and having to amend.

An hourly bank feed reconciled against ERP invoices. Runs on a six-hour overlapping window because banks backfill, and stays idempotent so the overlap costs nothing. Extracts tax IDs out of free-text bank descriptions and matches instalments and grouped payments.

My first project here, and the one that started the rest: an Excel VBA macro rebuilt as an R/Shiny app. Consolidates statements from multiple accounts whose headers never sit on the same row, and classifies every movement through 78 rules. 525 movements, none left unclassified.