Portfolio project · Python + AI (sample data)

From campaign data to decisions that raise ROAS

Transparency note: "EcoClean Hogar" is a fictional company used to build this Google Ads analytics and prediction platform end to end. It is not a real client; it is the same kind of technical work I apply to real accounts.

An analytics and prediction platform for Google Ads campaigns, built in Python with Machine Learning and 2D/3D visualization.

The problem it solves

Google Ads reports tell you what happened. They do not tell you what will happen. When analysis stays in a spreadsheet, there is always a leak:

• The inefficient campaign that keeps getting budget out of habit.
• The cross-analysis nobody does by hand: CTR vs. CPC vs. ROAS at the same time.
• The decision to scale spend based on gut feeling, not on a model trained on your own data.
• The dataset that grows (10K, 1M rows) and the Excel file that can no longer cope.

What the platform does: measure, see and predict

KPIs: executive console dashboard

dashboard.py calculates aggregated CTR, CPC, CPA, ROAS, cost, conversions and revenue, and offers a menu to launch any chart.

Visualization: 10 2D/3D charts

graficos3d.py: scatter3D, heatmap, funnel, treemap, sunburst and more, each one crossing variables that Excel cannot.

AI prediction: future conversions

modelo_ads.py trains a neural network (MLPRegressor) that predicts conversions from impressions, clicks, CTR, CPC and cost.

How it works: from a dataset to a prediction in 4 steps

1. Data: the dataset is loaded: 300, 10,000 or 1,000,000 rows, same columns, same code with no changes.
2. KPIs: metrics are aggregated (CTR, CPC, CPA, ROAS, cost, conversions, revenue) and shown in the executive dashboard.
3. Visualize: variables are crossed with 10 on-demand 2D/3D charts by campaign, channel, device, hour or date.
4. Predict: the model estimates conversions with a neural network trained on the client's real data, and reports R² and MSE.

Tech stack: everything runs locally, no cloud or external APIs

• Python 3 — the base language of the whole platform.
• Pandas / NumPy — loading, cleaning and aggregating datasets (300 to 1M rows).
• Matplotlib — static charts for the executive dashboard and reports.
• Plotly — the 10 interactive 2D/3D visualizations.
• Scikit-Learn — neural network (MLPRegressor) that predicts conversions.
• OpenPyXL — reads the .xlsx datasets without needing a database.

Results: the model tested at three scales

The 1,000,000-row dataset is sampled down to 50,000 to keep training fast, without touching the code.

How I work

• An architecture document that sets the goal, scope and rules before writing any code.
• Modular code: AI, charts and executive dashboard in separate files, each function with a single responsibility.
• Works the same with 300, 10,000 or 1,000,000 rows, tested at all three scales.
• 100% local: no database, no external API, no cloud. Full control over the client's data.