Continuum Lands: Real Estate Intelligence Orchestrated as a Graph

Problem

Deciding whether a property is worth investing in or developing usually means pulling together zoning rules, density coefficients, nearby services, comparable sales, and multi-year financial projections from several disconnected sources and specialists. That process is slow, inconsistent across analysts, and rarely delivered as a single, coherent document.

Approach

Continuum Lands turns an address into a professional report within a couple of minutes: land use, zoning, market trends, comparables, and, depending on the plan, financial projections or full due diligence. Behind that single request runs a longer, unattended sequence: geocoding and validation, data collection from several providers, financial modeling, narrative analysis, and rendering, before the report reaches the user.

Architecture

• The pipeline is modeled as a stateful graph rather than a linear script. After geocoding, data collection from multiple providers runs as parallel branches that rejoin before the report continues; later stages such as financial modeling, narrative analysis, and rendering run as distinct nodes with their own state, rather than one monolithic function trying to do everything in order.
• Financial analysis is computed once, deterministically, and kept separate from the narrative layer entirely: the numbers come from formulas, not from a language model reasoning about them.
• Every narrative section is grounded against one shared fact sheet built early in the pipeline, so figures and claims stay consistent across dozens of generated sections instead of drifting apart the way independently prompted sections tend to.
• Before any section reaches a language model, an embedding of the address and its collected data is checked against previously generated analyses for a close semantic match. Only a genuine miss spends a model call, so cost and latency scale with how novel a query actually is, not with total volume.
• Findings from earlier reports accumulate quietly by municipality, so an analysis in a region the system has already covered starts from more than a blank page.

Technical decisions

• API and execution are split: a request returns immediately with a job identifier while the pipeline runs in the background, since geocoding, several external data providers, a language model, and PDF rendering are all too slow to hold a request open for.
• Every external data source fails independently. If one is unavailable, the report’s completeness score drops but generation continues, rather than the whole analysis failing on a single unreliable dependency.
• Two rendering engines, one lightweight and one full-featured, share the same domain layer, fact sheet, editorial outline, and narrative composition, so different subscription tiers get proportionally deeper reports without duplicating the logic that decides what belongs in them.

Limitations

Coverage is currently limited to Mexican property markets, and analysis depth depends on how much data is available for a given address and its region.