Alejandro – Databricks, Pandas, Machine learning

A multi-tenant SaaS that helps restaurants track what their suppliers charge them. It connects to the restaurant's Outlook mailbox, picks out supplier invoices with an LLM, extracts every line item from PDFs and photos with a vision LLM, and sends a same-day alert when a supplier raises a unit price.

• Designed, built, and operated the multi-tenant SaaS alone in production with Python, Reflex, and PostgreSQL with row-level security.
• Built an invoice-ingestion pipeline that polled clients' Outlook mailboxes through Microsoft Graph, classified invoices with an LLM, and extracted line items from PDFs and photos with a vision LLM.
• Added a verification layer where the model only transcribed printed values and code checked each value against arithmetic relations learned per invoice layout, raising field-level accuracy from 96% to over 99%.
• Built a hand-labelled evaluation set of 100 invoices to measure extraction quality.
• Delivered same-day supplier price-increase alerts that surfaced an average of 260 EUR a month in previously unnoticed increases for the pilot client.

A demand forecasting and stocking optimization system for a global FMCG company with over $20B in revenue, used daily by supply planners to set stock levels across SKUs and geographies.

• Built demand forecasting models for more than 1,000 SKU and geography combinations.
• Developed a clustering model that segmented SKUs by seasonal behavior, with an LLM layer that chose the number of clusters and explained each segment's behavior.
• Designed and implemented the data-driven stocking strategy.
• Deployed the forecasts as a daily Databricks job that wrote results to the client's database.
• Built a Streamlit app on Databricks that let planners change inputs, monitor the model, and re-run forecasts.
• Reduced the number of planners needed from 8 to 1, cut stockouts by 10%, and cut time spent in the distribution center by 30%.

A residential solar subscription for an energy retailer. Households got solar panels and a battery for a flat monthly fee, while the retailer carried the variable energy cost, so the forecasting and optimization layer had to keep that cost as low as possible.

• Estimated rooftop solar potential from coordinates alone, factoring in area, tilt, and shading from nearby terrain and buildings using external solar and mapping APIs.
• Built forecasting models for wholesale energy prices, household energy consumption, and solar panel production.
• Built an optimization model that decided for each household when to charge or discharge the battery and when to buy from or sell to the grid, minimizing the retailer's total energy cost.

Credit and climate risk models for banks in a regulated environment that required explainable models, including climate stress testing of credit portfolios.

• Designed and implemented forecasting and risk models in Python, PySpark, Xarray, SAS, and SQL.
• Modelled climate risks on a bank's credit portfolio for the 2022 ECB Climate Stress Test.
• Defined new methodologies that fed forecasted climate risks into economic capital and pricing using panel data models.
• Forecasted long-term wildfire risk under NGFS scenarios using high-resolution geospatial satellite data.
• Explained model behavior with SHAP values and presented results to the client through dashboards and presentations.
• Led and trained a team of three junior consultants.