How we turn 50 pages of paperwork into an offer strategy.

Every prediction OfferWise makes leans on a labeled training corpus. We didn't scrape a generic real estate dataset — we built ours from the ground up, focused on the California buyer.

The corpus combines three data sources:

• Public-records crawls from 13 California municipal portals — permit history, code violations, NFIP flood claims, tax records. Refreshed on a rolling schedule.
• Inspection report findings labeled with category, severity, and repair cost ranges. These are extracted text findings — not scraped from third-party report providers — annotated against a fixed schema of 40 finding-category baskets.
• Disclosure-vs-inspection contradiction pairs — 1,324 hand-labeled cases where what the seller disclosed and what the inspector found don't agree. This is the dataset behind our contradiction model.

An OfferWise analysis is not a single model. It's a pipeline of specialized models, each doing one job well. The result is then synthesized into the buyer report.

• Category classifier. Sorts each finding into one of 40 baskets (roof, foundation, plumbing, electrical, etc.). Built on sentence-transformer embeddings + XGBoost.
• Severity ranker. Predicts how serious each finding is on a calibrated scale, accounting for inspector phrasing patterns (inspectors tend to underrate or overrate consistently).
• Repair-cost estimator. Predicts the dollar cost range for the finding, calibrated by region and home age. Outputs confidence intervals, not point estimates.
• Contradiction detector. Compares every disclosure claim against every inspection finding to surface mismatches.
• Public-records cross-reference. Pulls permits, code violations, NFIP claims, and assessor records for the property and matches them against the inspection and disclosure.
• Property similarity profile. Encodes the property as a high-dimensional vector so the system can find comparable homes in the corpus and ground predictions in actual comparable data.

Each model is trained separately and evaluated against held-out portions of the corpus. We don't ask a single language model to do everything — and we don't trust a single language model with any of it.

These are the numbers from our most recent training cycle, measured on held-out data:

The numbers we are most proud of, and the numbers we are honest about:

• Category accuracy at 85.8% means roughly 1 in 7 findings ends up in the wrong category at first pass. We surface these to the buyer with category-confidence indicators so they know which findings to double-check.
• Cost estimation MAE of $1,915 means our typical repair-cost estimate is within ~$2,000 of the true value. For a $30,000 roof, that's tight. For a $500 caulking fix, the error band is proportionally larger.
• Contradiction precision at 97.1% means when our model flags a disclosure-vs-inspection mismatch, it is almost always real. We tune this model to favor precision over recall — we'd rather miss some real contradictions than cry wolf.
• Severity accuracy at 65.4% is our weakest model. Severity is genuinely subjective — a "moderate" foundation crack means different things to different inspectors. We use this as a sanity check rather than a primary input to the offer math.

OfferWise is a strong tool for what it does. We try to be clear about what it doesn't do.

• We are not licensed inspectors. We don't replace a physical home inspection. We make the report you already have more useful.
• We are not lawyers, agents, or contractors. Our cost estimates and findings are inputs to your decision, not legal advice or contractor quotes.
• Vision-based image analysis is partial. When inspection PDFs require visual extraction (scanned reports, image-heavy documents), our system describes embedded photos in detail. Most modern text-extractable PDFs bypass this path — image analysis on every report is on the roadmap, not the current default.
• We are strongest in California. The corpus is California-focused. Out-of-state analyses run on the same models but with less locality-specific training data. National coverage is growing.
• Predictions, not certainty. Every output is a probability or a range. We try to be calibrated, not confident. Use our reports as the start of your due diligence, not the end of it.