SmartCAMA

Challenge

Property assessment data is available—but difficult to review at scale.

Philadelphia publishes extensive property, sales and assessment records. However, raw tables alone make it difficult to identify geographic patterns, compare modeled values with official assessments, or understand how an individual property’s value has changed over time.

The project focused on two audiences with different needs:

• Tax assessors need a citywide, parcel-level view for locating unusual assessment differences.

• Property owners need a simple way to search an address and understand assessment history.

Solution

One data platform, two user experiences.

SmartCAMA combines property records, historical assessments, parcel geometry and machine-learning predictions in a shared cloud data pipeline.

Tax Assessor Review

An interactive parcel map allows assessors to:

• Compare ML-predicted and official assessed values

• Switch between value and difference layers

• Filter properties by ZIP code

• Highlight parcels with large discrepancies

• Review citywide value distributions

Property Owner Widget

A public-facing lookup experience allows residents to:

• Search by address

• Review property characteristics

• Explore assessment history

• Identify significant year-over-year changes

• Compare land and building values

From Open Data to an Interactive Map

Building an end-to-end geospatial data pipeline

The system ingests Philadelphia property and parcel datasets, prepares them for analysis, trains a valuation model in BigQuery ML, and publishes the resulting information as lightweight browser-ready assets.

The pipeline separates data into raw, prepared, core and derived layers, making each transformation easier to trace and reproduce. Parcel geometry is stored as geographic data in BigQuery, exported as GeoJSON, and converted into vector tiles with GDAL/OGR for efficient map rendering.

Modeling Property Values

Turning recent residential sales into assessment estimates

A random-forest regression model was trained using recent arm’s-length residential sales. The preparation process excludes bundle transactions, nominal transfers and extreme price outliers that could distort the model.

Features include:

• Living and total area

• Bedrooms and bathrooms

• Building age and condition

• Building and property type

• ZIP code

• Sale recency

The resulting predictions are joined back to parcel geometry and compared with official tax-year assessments.

SmartCAMA is a prototype. Its predictions are intended for exploration and review, not as official property assessments.

Designing for Two Levels of Understanding

Citywide patterns and individual property stories

The assessor interface emphasizes comparison and anomaly detection. Color-coded parcel layers expose geographic patterns, while the discrepancy filter narrows attention to properties that may require further review.

The owner interface reduces the same underlying data into a more understandable narrative: what the property is, how its assessed value has changed, and how land and building values contribute to the total.

Outcome

A reusable prototype for transparent, parcel-scale assessment review

SmartCAMA demonstrates how cloud data engineering, machine learning and Web GIS can work together as a single public-sector data product.

The prototype delivers:

• A reproducible cloud-based data workflow

• Parcel-scale visualization for approximately 424,000 residential properties

• A shared data foundation for professional and public-facing tools

• Efficient browser rendering through vector tiles

• A more accessible way to explore assessment history

Reflection

What I learned

Reliable data foundations matter more than model complexity

The most difficult part of building SmartCAMA was not training the random-forest model—it was creating a consistent data foundation. Property records, assessment histories, and parcel geometries came from different sources and had to be standardized around a shared property identifier before they could be analyzed together. I learned that even a strong model becomes unreliable when joins, field types, or historical records are inconsistent.

Web-map performance begins in the data pipeline

Rendering hundreds of thousands of parcel polygons directly as GeoJSON would create an unnecessarily heavy browser experience. Converting the map-ready dataset into vector tiles allowed the application to load only the properties needed at each zoom level. This showed me that responsive Web GIS depends as much on data formats and publishing architecture as it does on front-end code.

Predictions need context, not just presentation

A modeled property value can look authoritative when displayed on a polished map. I learned that prototype predictions must be clearly separated from official assessments and accompanied by methodology, limitations, and uncertainty. Future development should include neighborhood-level residual analysis, fairness evaluation, model explanations, and user testing with both assessors and residents.