How ATourOf turns remodeling projects into comparable data.
Source provenance, contractor confirmation and confidence matter more than pretending every field can be inferred automatically.
1. Data layers
ATourOf separates raw contractor submissions, machine observations, contractor confirmations, normalized model features and public presentation. A value can move between those layers only when its source and permitted use are known.
2. Photo analysis
Vision models may identify visible attributes such as cabinet style, apparent material, layout cues or visible brands when evidence is strong. They must not infer exact project cost, concealed mechanical work, permits or structural scope from appearance alone. Ambiguous observations remain uncertain.
3. Normalization and eligibility
Project records track cost precision, quantities, inclusion flags, scope flags, confirmation status and a data-quality score. Low-quality or unconfirmed records can remain useful operationally while being excluded from model training or aggregated statistics.
4. Comparable projects
The initial estimating approach favors nearest comparable completed projects over opaque point predictions. Relevant dimensions can include geography, year, square footage, layout, layout change, cabinet tier and quantity, countertop area/material, appliance package and confirmed scope.
5. Confidence
Small or heterogeneous comp sets should widen the returned range and reduce confidence. ATourOf does not consider false precision a feature.
6. Public benchmark publication
Internal exploratory cohorts may exist with smaller samples. Public derived benchmarks require a sufficient comparable sample and privacy review; the current planned minimum is 10 qualifying projects, with higher thresholds where appropriate.
7. Cost visibility
A contractor can provide exact cost for private model use while allowing only a range—or nothing—to appear publicly. A public project page does not imply that an exact contractor cost is available behind it.
8. Updates
Methodology will evolve as the verified corpus grows. Material changes to normalization, thresholds or benchmark logic should be reflected on this page rather than hidden behind model changes.