Jose Tomas Lavin Leon

An ML observability system I designed and led the development of to monitor the production credit scoring models that decide loan approvals for Aplazame, a BNPL fintech. The system is built around three layers: service health (Sentry alerting on infrastructure and pipeline failures), data quality (Airflow-orchestrated checks for drift across incoming feature distributions using PSI, KS, and Wasserstein distance), and model performance (also orchestrated via Airflow, feeding a live Dash dashboard that the risk and data science team uses to track how each model is behaving in production). A separate monthly job computes SHAP values on ephemeral EKS pods and writes the results to S3, so feature-level explanations are available on demand without increasing the cost of daily inference.

Before this, there was no automated way to catch model degradation. Issues surfaced only when someone noticed a problem downstream. Now, data quality issues are flagged within minutes, model performance is reviewed monthly, and SHAP explanations are ready for any model version, which has also sped up internal audits. It's become the reference monitoring architecture that Aplazame expects new models to ship with.

An MCP connection I built for Panel Ciudadano that lets the team query their survey and operational data directly through AI agents, instead of writing SQL or waiting on a data team. It sits on top of a GCP-native data pipeline I also designed, which ingests data from a PostgreSQL source and transforms it with dbt into BigQuery. The dbt models double as a semantic layer, defining business metrics like completion rates consistently in one place, so every AI agent query resolves against the same definitions instead of ad-hoc SQL that could interpret a metric differently each time.

Panel Ciudadano is both the client and the primary user its internal team needed a way to get answers from survey data without writing SQL or routing every question through a data request queue. I made the core architecture decisions myself, from the pipeline and semantic layer design to the structure of the MCP layer. The platform is currently in its improvement phase, with the goal of letting anyone at the company ask a question like "what's our completion rate for survey X" in natural language and receive a consistent, accurate answer instead of waiting on a ticket.

I led the design and implementation of a segmented credit-scoring model for Aplazame, a BNPL fintech, to approve or decline loans in real time at the point of purchase. It uses graph-based neighborhood features that capture relationships between applications and is served via an API that handles thousands of predictions daily. I also implemented MLOps frameworks like MLflow to track experiments and manage the model lifecycle, and we use A/B testing to validate and roll out new models before full deployment. I also contributed to several other credit-scoring models on Aplazame's risk team, using gradient-boosting techniques such as XGBoost and CatBoost.

The client is Aplazame itself, and the end users are the risk team, who rely on these models to approve or decline transactions in real time, and, indirectly, Aplazame's retail customers, whose checkout experience depends on fast, accurate decisions.

Across the models I contributed to, the combined impact was a 25% annual reduction in portfolio risk while maintaining the existing acceptance rate, meaning fewer bad loans approved without declining more good customers, alongside models that are simpler to maintain and retrain than the legacy monolithic approach.