Property Valuation That's
Accurate to 97.2% in Under a Second.
Agix partnered with HouseCanary to build a next-generation Automated Valuation Model, combining a 100M-property data graph, computer vision condition scoring, and multi-signal neighborhood AI, that delivers lender-grade valuations 300× faster than a traditional appraisal at 1% of the cost.
The data company that put lender-grade property intelligence into a single API call.
HouseCanary is a San Francisco–based real estate data and analytics company founded in 2013 with a singular mission: give anyone making a decision about a property access to the same quality of analysis that top institutional lenders use. Their platform covers 100M+ residential properties across all 50 US states, delivering automated valuations, market forecasts, and risk signals to mortgage lenders, institutional investors, insurance carriers, and iBuyers. When HouseCanary approached Agix, their core AVM was producing solid results, but the gap between their accuracy ceiling and what the best institutional underwriters needed for full automation remained stubbornly wide. The mandate: close that gap permanently, at scale, in real time.
89% accuracy is fine for browsing. For a $600,000 lending decision, it's the difference between a portfolio and a loss.
Traditional appraisals take 3–5 days and cost $400–800 per property. Lenders needed a valuation engine that matched the judgment of an experienced human appraiser, in a single API call.
Structured data couldn't see condition or micro-neighborhood quality
MLS records and deed data told the model a property had 3 bedrooms and 1,800 sq ft, but couldn't tell it the roof was failing, the block had three vacant lots, or that a development two streets over was driving appreciation. Condition and context signals account for 15–25% of fair market value. Without computer vision, the AVM was systematically mismeasuring an enormous slice of its universe.
Comparable selection was rules-based, and rules can't read the market
Find homes within X miles, last Y months, similar beds and sq ft, defensible in stable markets, but consistently wrong in urban neighborhoods with high heterogeneity, rural markets with sparse volume, or rapidly appreciating corridors where six-month-old sales are already stale. An expert appraiser weighs location micro-factors, renovation quality, and time-adjusted trajectory. A filter form can't.
Point estimates without confidence intervals left lenders unable to automate
A loan on a property valued at $445,500 ± $4,500 is a fundamentally different risk profile than one valued at $445,500 ± $68,000. Without quantified uncertainty, underwriting teams couldn't set automated approval thresholds, and every borderline case required human review, defeating the purpose of automation entirely. Calibrated confidence intervals were as critical as accuracy itself.
Six AI stages, from raw property data to a lender-grade valuation in under one second.
Data Sources → Ingestion → Valuation AI → Market Forecasting → Risk Scoring → Investor & Lender Decisions, running for every property, in every market, in real time.
Six AI systems that turned a data platform into the most accurate AVM in the US residential market.
Each component was engineered to work independently and as part of a unified valuation pipeline, sharing a single property data graph, updating continuously as new sales and imagery arrive, and improving with every transaction the model observes.
100M Property Data Graph
The foundation of the entire valuation system: a continuously updated graph database linking every US residential property to its full transactional history, structural characteristics, ownership chain, tax assessment record, lien history, and neighborhood context. Built on a multi-source ingestion pipeline normalizing data from 3,100+ county assessor databases, MLS feeds across all 50 states, deed and lien records, and permit databases, updated within 24 hours of any new transaction anywhere in the coverage universe. The graph structure allows the valuation models to traverse relationships that flat databases can't represent: understanding how a specific property connects to its micro-neighborhood, how that neighborhood connects to its broader market, and how that market is trending relative to comparable metros. With 100M+ nodes and billions of edges, the data graph is the competitive moat that makes everything else possible.
Computer Vision Condition Assessment
A custom computer vision pipeline that processes street-level and aerial imagery for every property in the coverage universe, extracting 47 condition and quality signals from each image set: roof condition, facade material and quality, landscaping maintenance, visible deferred maintenance, driveway and hardscaping quality, window and door condition, and structural anomaly flags. The model was trained on a proprietary dataset of 2.4M labeled property images paired with known sale prices and appraiser condition ratings, enabling it to learn the precise relationship between visual condition signals and market value impact. Running on demand at inference time, the vision pipeline adds a condition-adjusted value modifier to the base statistical model, correcting for the systematic errors that structured data alone consistently makes on properties at the tails of the condition distribution.
Neighborhood Trend Vectorization
A spatial AI system that constructs a real-time "neighborhood vector" for every block in the coverage universe, encoding the current trajectory of 18 hyper-local market signals: price per sq ft trends, days-on-market velocity, list-to-sale price ratio, new construction activity, permit volume, vacancy rate, school rating trends, walkability score, transit access, crime trend, demographic shift signals, and commercial development pipeline. These neighborhood vectors feed directly into the comparable selection engine and the confidence interval model, allowing the AVM to automatically down-weight comps from micro-markets that are trending differently from the subject property's block, and to widen confidence intervals in markets experiencing rapid change or high heterogeneity. The vectorization model updates daily as new transactions, permits, and listing activity arrive, ensuring the neighborhood context is always current at inference time.
Comparable Selection Engine
A learned similarity model that selects and weights comparable sales the way an expert appraiser would, not by applying fixed distance and recency filters, but by understanding multi-dimensional property similarity across location quality, physical characteristics, condition, and market trajectory. The engine generates a similarity score for every potential comp in the coverage universe and selects the optimal set for each valuation, typically 12–20 comparables weighted by relevance rather than the traditional fixed set of 3–6. It automatically handles edge cases that defeat rule-based systems: sparse rural markets (by expanding the comp search radius while adjusting for location premium differentials), heterogeneous urban blocks (by weighting comps from the most similar micro-context), and rapidly appreciating markets (by applying time-trend adjustments that bring older comps forward to current market conditions rather than simply excluding them).
Confidence Interval Generation
A calibrated uncertainty quantification layer that produces a defensible valuation range, not just a point estimate, for every AVM output. The model was trained to accurately predict its own error: using a conformal prediction framework, it generates intervals that contain the true sale price at the specified confidence level (e.g., 90% intervals contain the true value in 90% of cases when tested on held-out data). The width of the interval reflects the genuine difficulty of the valuation: a well-maintained home in a stable, transaction-rich neighborhood with identical recent comparables gets a tight interval; a renovated Victorian on a block with highly heterogeneous housing stock in a market with sparse recent sales gets a wide one. Lenders can set automated approval thresholds by confidence interval width, approving loans where the AVM is highly certain, flagging those where it isn't, and eliminating human review for the majority while concentrating underwriter attention on the cases that genuinely warrant it.
12-Month Price Forecast Model
A forward-looking price appreciation model that generates a 12-month price forecast and a 5-year appreciation outlook for every property and market in the coverage universe, using a multi-scale temporal model that combines macro economic indicators (Fed rate trajectory, housing supply pipeline, employment growth), metro-level supply/demand dynamics (inventory levels, permit volume, absorption rate), and block-level trend signals from the neighborhood vectorization system. The forecast outputs a point estimate and a probabilistic range, enabling lenders to model collateral risk over the life of a loan, investors to identify markets where current prices undervalue future appreciation potential, and insurance carriers to adjust coverage based on projected value trajectories. With a 6.4% MAPE on 12-month price forecasts versus the industry average of 12.8%, the model outperformed every competing product in independent validation testing.
What happened when every property decision was powered by AI.
Measured across HouseCanary's deployed AVM platform, spanning institutional lenders, investors, and insurance carriers using the full AI valuation stack in production.
Valuation accuracy on independent held-out test sets, 8.2 percentage points above the 89% industry standard for best-in-class AVMs, and within range of human appraisers on comparable properties
Full AVM response time including computer vision condition scoring, comparable selection, confidence interval generation, and 12-month price forecast, delivered in a single API call at 300× the speed of a traditional appraisal
Mean Absolute Percentage Error on 12-month price forecasts, half the 12.8% industry average, enabling confident collateral risk modeling over the life of a loan with materially narrower uncertainty bands
Accuracy identifying high-risk valuations that warrant human review, enabling lenders to automatically approve the low-risk majority while concentrating underwriter attention on cases that genuinely need it
Active US residential properties with full AVM, confidence interval, and 12-month price forecast available on-demand, the most comprehensive single-source coverage of the US residential market
Total loan volume underwritten by institutional lenders using HouseCanary as the primary valuation source, at 1% of the time and cost of traditional appraisal-based underwriting workflows
We've underwritten more than $4 billion in loans using HouseCanary as our primary valuation source. The accuracy is indistinguishable from a human appraisal on the properties where we've run both, and it runs in under a second at 1% of the time and cost. What Agix built here isn't a faster version of what we had before. It's a fundamentally different capability that has changed how we think about the economics of underwriting.
Three architectural decisions that made this the most accurate AVM in the US residential market.
Multi-modal data fusion, not another regression model
The core architectural insight was that property valuation is a multi-modal problem, it requires integrating structured transactional data, unstructured visual data, and spatial time-series data into a single coherent prediction. Most AVMs are fundamentally regression models with more features. We built a model ensemble that treats each data modality as a first-class input: a gradient-boosted tree model on structured property and transaction features, a convolutional network on visual condition data, and a spatial transformer on neighborhood vector embeddings, with a learned fusion layer that weights each modality's contribution based on how much information it adds for each specific property type and market context. The fusion approach is why accuracy improved so dramatically: each modality fills gaps the others leave, rather than all three models struggling with the same information deficiencies.
Calibrated uncertainty as a first-class output
Most AVM vendors treat confidence intervals as a marketing feature, a wide range bolted on as a disclaimer rather than a rigorously calibrated statistical output. We treated confidence calibration as a core accuracy requirement equal in importance to the point estimate itself. The model is evaluated not just on how close its point estimates are to true sale prices, but on whether its stated confidence intervals contain the true value at the promised frequency. A model that says it's 90% confident should be right 90% of the time on held-out data, not 70%, not 95%. This calibration discipline meant that lenders could actually rely on interval width as an automated underwriting signal, rather than treating it as a rough directional indicator. The value of a precisely calibrated confidence interval, in enabling automated approval at scale, turned out to be as commercially significant as the improvement in raw accuracy.
Continuous learning from every new transaction
An AVM's accuracy at deployment is not its accuracy six months later, unless it learns from every new transaction it observes. We built the model update pipeline as a core system component, not an afterthought: every sale in the coverage universe that closes is automatically ingested, the model's prediction for that property is retrieved from the prediction log, the error is computed, and the result feeds into a continuous fine-tuning cycle that runs weekly for the comparable selection and neighborhood vectorization components and monthly for the full ensemble. Markets that are changing fastest get the highest update frequency. The result is a model whose accuracy improves rather than decays over time, and whose neighborhood context is always current, never representing a market that existed six months ago and has since moved.
From valuation accuracy to investment intelligence, the full platform, live.
AVM accuracy enables downstream intelligence: Investor Opportunity Finder, Portfolio Risk Monitor, and Market Heat Maps, all powered by the same data graph and valuation engine.
What powers this system.
Automated Valuation Models (AVM)
Multi-modal property valuation combining structured transaction data, computer vision condition scoring, and neighborhood vector embeddings, delivering lender-grade accuracy at API scale.
Computer Vision for Real Estate
Custom vision pipelines trained on millions of labeled property images, extracting 47 condition and quality signals from street-level and aerial imagery to correct systematic errors in data-only valuation models.
Market Forecasting & Price Intelligence
Multi-scale temporal models generating 12-month price forecasts and 5-year appreciation outlooks, combining macro economic signals, metro supply/demand dynamics, and block-level trend vectors for 6.4% MAPE accuracy.
Property Data Graph Infrastructure
Graph databases normalizing 3,100+ county assessor feeds, multi-state MLS data, deed and lien records, and permit databases, updated within 24 hours of any new transaction anywhere in the coverage universe.
Calibrated Uncertainty Quantification
Conformal prediction frameworks generating defensible confidence intervals that contain the true value at the stated frequency, enabling automated underwriting thresholds with 94.7% risk detection accuracy.
Fintech & Lending AI Platforms
End-to-end AI systems for mortgage lenders, iBuyers, institutional investors, and insurance carriers, from automated underwriting workflows to portfolio risk monitoring and market opportunity identification.
Common questions about building lender-grade property AI at scale.
The comparable selection engine was specifically designed for sparse transaction environments. For rural markets or highly unique properties where recent local sales are unavailable, the engine expands its search radius progressively while applying a learned location-premium differential, adjusting comp prices for the distance-to-subject premium or discount based on market topology rather than applying a flat geographic filter. In the most sparse markets, the model shifts weight toward the neighborhood trend vectorization and the computer vision condition signals, which are available for every property regardless of local transaction volume. The result is that accuracy degrades gracefully in sparse markets rather than collapsing, the model correctly widens its confidence interval to reflect genuine uncertainty rather than producing a spuriously precise estimate based on distant or dated comparables.
The vision pipeline integrates imagery from multiple sources, including Google Street View, Nearmap aerial updates, and county assessor photo archives, with different update frequencies. Street-level imagery typically refreshes on a 1–3 year cycle for most US markets; aerial imagery refreshes annually for high-coverage areas. The model tracks imagery age as a feature and automatically increases the weight of structural and tax record data (which captures major renovations through permit records) when imagery is more than 18 months old for a given property. Permit records, which capture roof replacements, additions, and major renovations, are integrated within 24 hours of filing and update the condition prior directly, so a property with a recently permitted roof replacement gets an updated condition signal before new imagery is available. The system is designed to be transparent about imagery age in its API response, so lenders can apply their own policy on how to treat valuations with older visual data.
Non-standard property types, mixed-use residential/commercial, properties with accessory dwelling units, multi-family properties up to 4 units, are covered with specialized model variants rather than the core single-family AVM. Each variant is trained on a property-type-specific dataset and uses a tailored feature set: for ADU-equipped properties, the income-potential contribution of the unit is explicitly modeled using local rental market data; for mixed-use properties, the commercial component is valued using a separate income-capitalization approach blended with the residential comparable analysis. The confidence intervals are typically wider on non-standard types reflecting genuinely higher valuation uncertainty, and the API response flags the property type classification so downstream systems can apply appropriate review policies. Properties the model cannot classify confidently are flagged for human appraisal rather than attempting an uncertain automated estimate.
Yes, the platform was built to serve multiple downstream use cases from a single valuation infrastructure. For property insurance carriers, the computer vision condition model is particularly valuable: identifying properties with deferred maintenance, aging roofs, or structural anomalies that represent elevated claims risk, enabling risk-based premium pricing and underwriting decisions rather than blanket coverage. For iBuyers, the price forecast model and the investor opportunity score (combining appreciation potential, rental demand, and risk signals) directly feed acquisition decision workflows. Each use case has a tailored API response schema that surfaces the signals most relevant to that decision type, a lender gets confidence intervals and UCDP-compatible outputs, while an iBuyer gets IRR projections and neighborhood trajectory scores. The core valuation engine is the same; the downstream intelligence layer adapts to the decision context.
Fair lending compliance was a first-class design requirement. The model explicitly excludes protected class demographic features, race, ethnicity, religion, national origin, gender, disability status, from all inputs. The neighborhood vectorization system uses economic and physical signals (employment, income, transit access, school ratings) rather than demographic composition, and the model is regularly tested for disparate impact using independent fair lending audits. The comparable selection engine is tested to ensure it doesn't produce systematically lower valuations in majority-minority neighborhoods when property and market characteristics are controlled for. Audit logs of every valuation decision, including the comparable set used, the condition signal applied, and the confidence interval rationale, are retained and available for regulatory examination. Lenders using the platform receive a compliance documentation package that supports ECOA adverse action notice requirements for any valuation-based credit decision.
Ready to build a property valuation platform accurate enough to underwrite at scale?
Most projects go from kickoff to deployed AI system in 8–16 weeks. Let's talk about what a lender-grade AVM, risk scoring engine, and market forecasting platform could do for your underwriting workflows.