baobab AI (formerly baobab soluciones), Madrid
Forward-deployed at a leading multinational brewer (2024 – 2026): lead in-house data scientist and project manager, owning each project from business definition to production and hypercare.
- Distributor next-best-action alerts: owned end to end the engagement’s highest-value product, daily alerts for independent wholesalers’ sales reps built on sell-out data. Agreed the API contract with the offshore integration team and each distributor, and fed the reps’ answers back from their ERPs to silence or re-raise alerts.
- Churn prediction: replaced an external vendor’s model with an in-house one scoring tens of thousands of customers monthly (LightGBM, calibrated, SHAP reasons): 2–3× the alert precision, and flagged customers churn 7–10× more often than average.
- Sales alerting platform: led the platform that turns sales, agreement, equipment, churn and P&L data into thousands of monthly alerts pushed to Salesforce.
Other client projects:
- Airport assistance demand: forecast reduced-mobility passenger assistance at airports (LightGBM, MSTL, Prophet, web-scraped covariates), deployed with Docker.
- Solar data quality (project manager): rule-based anomaly detection plus autoencoder reconstruction of solar measurement data, shipped as v1.0 with CI and SonarQube.
- Social services demand (project manager): forecast demand for a national disability-support organisation’s services (financial aid, private tutoring, guide dogs) to improve how resources are allocated.
Platform, generative AI and pre-sales:
- Rolled out MLflow company-wide as an internal service with Google SSO and group access control, plus a starter project template.
- Built baobab’s time series framework: backtesting, leakage-checked features and per-series model routing (Kedro, darts, polars).
- Shipped generative AI marketing tools end to end, including a production lead-generation assistant for an energy campaign that turns a conversation into a tailored solution summary, with lead analytics for marketing.
- Pre-sales: first client meetings, scoping, proposals, mockups, presentations and demos that help close deals.
- Agentic development with Claude Code and a custom multi-agent workflow, to iterate with the business much faster.
- Credit risk: default-probability models for financial guarantees, and loan default forecasting so a lender could start collection and legal processes early.
- Computer vision: classified electrical grid assets from aerial imagery (CNNs, YOLO, GroundingDINO).
- Quality analytics: explained defects in electrical transformer manufacturing with CatBoost and SHAP.
- Forecasting and causal inference: demand for new retail products and accessory stock; causal impact of campaigns on beverage sales.