After learning to train models at Lambda School, I wanted to see what AutoML changed—and what it did not. DataRobot's inaugural academy cohort let me test that question hands-on. The clips below show the modeling work, not automation presented as magic.
As a member of the inaugural DataRobot 10x Data Science Academy cohort, I used DataRobot's automated machine learning (AutoML) platform to move faster from exploratory analysis to deployment—without giving up explainability.
DataRobot Platform (AutoML, model blueprints, validation, What-If / deployment dashboards)
Three builds: Stack Overflow developer salaries, vehicle fuel efficiency (MPG), and county-level COVID spread. Each started with a domain question, then worked through feature importance, blueprints, lift curves, and dashboards.
Interpreting models (not just accuracy), comparing blueprints, and using DataRobot's visuals — feature effects, correlation maps, and deployed What-If tools — to explain predictions to a stakeholder.
Developer role type emerged as a strong salary driver, while the blueprint showed how extensive text features were numerically encoded during prep. The validation plots told the more useful story: the model tended to over-predict higher salaries and under-predict lower ones. Lift and predicted-vs-actual views made that pattern visible, and the feature-importance map surfaced correlation clusters among the colored features—especially oranges and greens. The built-in What-If dashboard then turned the trained model into an interactive salary conversation.
The deployment path did more than support exploration. It gave stakeholders a way to adjust inputs, compare scenarios, and keep the model’s behavior inspectable.
This workflow predicts miles-per-gallon from mechanical specs. As manufacturers move to EVs, a miles-per-kWh model would lean on physics—weight and aerodynamics—rather than cylinders, transmissions, or ethanol-heavy fields from classic MPG datasets. Many of those design questions can be explored in simulation before hardware exists.
ROC, coefficients, blueprints, and lift views gave the county-infection model more than one angle of inspection. Multicollinearity does not always hurt accuracy, but it can make individual effects harder to explain; the correlation map below shows twelve distinct clusters. Feature-effect, gain, and cumulative-lift views completed the validation story.
Multicollinearity does not necessarily reduce accuracy. It can, however, blur how confidently you attribute an effect to one correlated input—worth checking before presenting coefficients to anyone else.
Feature correlation map: twelve correlation clusters across inputs.
