Ottawa Home Atlas project infographic A technical project dossier covering the system's Census GIS lineage, catalogue scale, data reduction, geographic layers, autonomous daily pipeline, predictive-model validation, production safeguards, human direction, and toolchain. PROJECT DOSSIER · JULY 2026 OTTAWA HOME ATLAS A regional property-research system combining current listings, physical access, public demographics, elections, safety context, market history, and predictive models. Built as a local GIS workshop, distilled into a lean public service, then taught to fetch its own listings and rebuild its estimates every day. THE LIVE CATALOGUE 5,801 ACTIVE LISTINGS + 660 recently removed 25 MUNICIPALITIES within a 120 km region 4,996 VALUE ESTIMATES 805 estimates withheld 2,658 HEATMAP CELLS location value surface FROM WORKSHOP TO RUNTIME DEVELOPMENT CATALOGUES ≈ 3.8 GB Census · boundaries · crime · immigration elections · water · listings · derived models DISTILLED PUBLIC SERVING CATALOGUE 19 MB 6,461 market records · 239 geographic contexts · 1,571 polling contexts ≈ 99.5% smaller · public web process needs only DuckDB, FastAPI, and Uvicorn ONE PROPERTY, MANY CONTEXTS MARKET price · beds · baths · area new / removed lifecycle ACCESS nearest hospital · downtown water proximity · waterfront PEOPLE age · language · income 10-year population scenario PLACE crime context · federal poll candidate vote breakdown 28,762 MAPPED WATERBODIES 26 ONTARIO HOSPITALS 1,571 POLLING CONTEXTS THE AUTONOMOUS DAILY FARM Runs around 04:30 Toronto time · fetches, trains, validates, and switches without human intervention 1 FETCH 5 public listingsources 2 GUARD reject unsafesnapshots 3 ENRICH deduplicate andassign context 4 TRAIN CatBoost + spatialvalidation 5 ACTIVATE atomic snapshotswitch FIRST PRODUCTION FARM 6m 06s wall time 449 MB peak memory 6 new records found THE MODEL HAD TO EARN ITS PLACE Median absolute percentage error · lower is better · 20 km blocked geography Ridge baseline 15.5% CatBoost structural model · selected 13.4% CatBoost + spatial residual · rejected 14.5% FORWARD-TIME CHECK 10.5% median error on 428 newly listed properties PREDICTION DISCIPLINE 5-fold out-of-fold No displayed listing trains its own estimate. Weak support means the estimate is withheld. FAIL CLOSED, NOT LOUD Transactional crawl · 50% inventory floor · staged model Validated catalogue · atomic switch A blocked source, broken parser, failed model, or incomplete snapshot leaves the live site untouched. THE FULL GIS LINEAGE Home Atlas grew from the original Statistics Canada tract explorer—not from a blank map. Census geography grew into crime, elections, and demographics; listings made that foundation a daily property system. 127 WIZARD MESSAGES full project lineage 3,273 WORDS OF DIRECTION genuine owner prompts 15 MEDIAN WORDS mean: 25.8 190 LONGEST MESSAGE shortest: 1 word MESSAGE LENGTH · WORDS Equal 25-word bins · most decisions arrived as short, iterative instructions 82 1–25 25 26–50 10 51–75 5 76–100 4 101–125 0 126–150 0 151–175 1 176–200 TOOLS & EFFORT SYSTEM TOOLCHAIN DuckDB · Python · TypeScript MapLibre GL · PMTiles · FastAPI CatBoost · scikit-learn Playwright · Caddy · systemd Research, GIS, modelling, testing, and operations COLLABORATIVE BUILD 12 calendar days The Wizard set goals, constraints, UX direction, and phase approvals through 127 messages. Apprentice researched, coded, tested, modelled, packaged, and deployed the resulting system. Build ledger: 5,542 authored source lines across 26 application, test, and deployment files; 21 automated tests. Transcript scope: founding Statistics Canada tract request through the initial Home Atlas dossier; genuine owner-authored messages only. Automatic context/memory prompts and later counting requests excluded. Whitespace-delimited words; equal 25-word bins. Prices are asking-price models, not appraisals. This infographic may contain AI slop.