Notes on file
Blog
2026-11-16
Field notes: the afternoon we almost shipped a feature that just reproduced zip codeA field-notes account of catching feature leakage: a "neighborhood quality" variable that turned out to just be zip code in disguise.
2026-11-16
What is a kickout rate in AVM performance monitoring?A plain explanation of kickout rate in AVM monitoring, how it ties to confidence score thresholds, and why fallback to appraisal climbs on certain parcels.
2026-11-09
How to generate adverse action reason codes from a learned imagery embedding in an AVMA data-science walkthrough for turning an opaque image embedding feature in an AVM into ECOA-compliant adverse action reason codes.
2026-11-09
Proxy discrimination and fair lending review for imagery featuresWhat proxy discrimination means for AVM features, and how to run a fair lending review before a neighborhood imagery input goes into production.
2026-11-02
Hedonic pricing model vs repeat-sales index: which one actually drives an AVMHedonic and repeat-sales models solve different problems. Here's which one actually underpins AVM point estimates, and where imagery fits.
2026-11-02
The refresh nobody lines up in the budgetA field-notes look at what changes on a parcel in twelve months, and why AVM teams underbudget the imagery refresh that catches it.
2026-10-26
Desktop appraisal vs. hybrid appraisal: what each one actually inspectsA field-by-field breakdown of what desktop and hybrid appraisals inspect, and why neither fills the curb-appeal gap in your model.
2026-10-26
What is UAD, and how do condition fields get populated?UAD standardizes appraisal fields like C1-C6 condition ratings, but those values still come from one appraiser's eyes. Here's where the gap sits.
2026-10-19
Field notes: the comp that matched on every field and looked nothing alike from aboveA comp matched every tabular field and nothing on the ground. Curb-appeal and context signal belongs inside the model, as a direct input.
2026-10-19
FSD in an AVM: what moves it when you add a featureWhat forecast standard deviation measures in an AVM, why adding a feature doesn't automatically tighten it, and how to test lift before you ship.
2026-10-14
Backtesting a curb-appeal signal against repeat-sales pairsHow to validate a new curb-appeal or parcel-context feature using repeat-sales price pairs before it goes into a production AVM.
2026-10-14
What GSD do you need to score curb appeal from satellite imagery?A data scientist's answer to the GSD question: what resolution actually resolves curb-appeal and parcel-condition signal for an AVM.
2026-10-13
How to join a per-parcel embedding vector to your AVM's record IDsA practical walkthrough of joining a per-parcel embedding file to your AVM's record IDs, including key normalization and refresh handling.
2026-10-13
What counts as a deferred-maintenance signal in curb-appeal imageryA working definition of deferred-maintenance signal in property imagery, what shows up at sub-0.5m resolution, and why tabular fields miss it.
2026-10-07
Adding a new data source to an AVM without a full retrainA practical path for folding a new data source into a live AVM: feature store conventions, shadow scoring, and when a full retrain is warranted.
2026-10-07
How to validate a new AVM feature without appraiser labelsAppraiser sign-off is slow and expensive. Here's how data science teams validate a new AVM feature using sale prices and residuals instead.
2026-10-07
Satellite imagery vs. street-level photos for property condition signalA practical comparison of overhead imagery and street-level photos as property condition inputs for AVM pipelines, and where each one breaks down.
2026-10-07
What is curb appeal scoring for automated valuation models?A plain-language look at curb appeal scoring in AVMs: what it measures, why tabular fields miss it, and how it slots into a valuation model.