How larsiusprime/openavmkit works
Project type and purpose
• OpenAVMKit is a Python library and toolkit for mass appraisal of real estate, providing data cleaning, enrichment, modeling, inference, and report generation.
• It ships as a PyPI package, includes CLI-style pipelines, Jupyter notebooks for interactive workflows, and connectors to cloud storage/APIs.
File structure
• Top-level:
– openavmkit/ ← core Python package
– notebooks/ ← example pipelines & interactive notebooks
– docs/ ← user and API documentation
– docker-compose.yml, multistage-dockerfile, entrypoint.sh ← containerization
– tests/ ← unit and integration tests
• Inside openavmkit/:
– core modules: cleaning.py, modeling.py, inference.py, pipeline.py, etc.
– cloud connectors: cloud/azure.py, huggingface.py, sftp.py
– synthetic data: synthetic/…
– studies & analyses: horizontal_equity_study.py, ratio_study.py, etc.
– resources: JSON settings templates, report markdown/CSS
– utilities: reusable helpers (census API, plotting, geometry)
README for technical stack
• Pure Python (>=3.x)
• Optional cloud APIs: Azure, SFTP, Hugging Face, US Census, OpenStreetMap
• Jupyter notebooks for examples
• Dockerized for reproducible environments
Main components & their interactions
a. Data Inputs
– Raw property records (CSV, database dumps)
– Census API calls
– OpenStreetMap enrichment
– Synthetic data generator (for testing/benchmarks)
b. Data Cleaning & Enrichment
– openavmkit.cleaning & cleaning pipeline notebooks
– Filters, data dictionary lookups, geometry routines
c. Feature Engineering & Modeling
– openavmkit.calculations, modeling.py, tuning.py
– Training statistical or ML models
d. Inference & Projection
– openavmkit.inference & projection.py
– Batch scoring of properties
e. Quality Control & Studies
– openavmkit.quality_control, horizontal_equity_study, ratio_study, sales_scrutiny_study
f. Report Generation
– openavmkit.reports → Markdown → HTML/PDF via CSS in resources/reports/
g. Cloud Storage & Deployment
– openavmkit.cloud.* interfaces push/pull data and artifacts
– Docker container bundles the runtime, entrypoint.sh orchestrates pipelines
h. User Interfaces
– CLI/pipeline scripts invoked from entrypoint.sh or directly
– Jupyter notebooks for exploratory use
Architectural patterns & principles
• Modular design: each stage (cleaning, modeling, reporting) is its own module
• Inversion of control: pipeline.py orchestrates modules via configuration JSON
• Extensibility: cloud connectors and synthetic data plug into base classes
• Reusability: utilities shared across modules