Case studySenior Analytics and Reporting Analyst
BI the business actually uses — Banco Pichincha (2023)
Five months in banking to close the distance between data and decision: dashboards the business actually uses, models in production and people able to sustain them.
- BI
- ETL
- scikit-learn
By the numbers
50+
users adopted the dashboards
+25%
in decision-making
>90%
accuracy of models in production
12
professionals trained in analytics
Context
The bank's analytics and reporting area produced dashboards the business never quite adopted, with slow ETL processes and models that stayed out of production.
The challenge
Turn BI into a tool the business uses daily and take predictive models to production with reliable accuracy.
How I did it
Designing from the decision
I led a five-person team building decision-oriented dashboards that more than 50 business users adopted. Each dashboard started from the decision it had to support and offered a path from the aggregate to the evidence, built in the semantic model rather than on the page.
One definition per concept
I centralized the measures in DAX and optimized the semantic models with DAX Studio and Tabular Editor. A measure defined once is the technical form of a business decision: delinquency or churn is calculated in one single way.
An ETL you can understand
I optimized the Power Query ETL by separating connection, preparation and model tables, with early row and column selection and validations before the model. The −35% in analysis times came from structural decisions, not from a performance trick.
Churn, delinquency and risk in production
I trained and deployed churn, delinquency and risk models with scikit-learn, evaluated by segment and by what each error costs: calling a client who was not leaving costs a call; not calling the one who leaves costs the client.
Training and governance
I designed a training program for 12 professionals —Power Query, modeling, DAX and communication, on their own data— and co-led data governance. The boundary was clear: exploring was free; redefining what delinquency is was not.
Impact
- +25% in decision-making on the adopted dashboards.
- −35% in analysis times from the ETL optimization.
- >90% accuracy and +35% in predictions from the models in production.
- +20% productivity in preparing and using information after the training program.
The sophistication of a solution does not by itself determine its value. What matters is the function it fulfills within a decision.