Every Building Is a Balance Sheet: Spatial Memory for Real-World Assets — Jordi Spranger, Spranic

A mortgage or insurance record can be updated every quarter while the building underneath it goes years without a fresh inspection. Jordi Spranger, founder of Spranic, opens this collaboration with Neo4j around that mismatch. His illustrative timeline starts with an assessment when a financial instrument is created, followed by another around year five. Between those visits, financial reporting continues, but knowledge of the property may remain scattered across photographs and PDFs.

The scale makes the mismatch consequential. Spranger puts global real estate at roughly $400 trillion and focuses on commercial property, including around $4 trillion in US commercial mortgages. These are market-context figures in the presentation, rather than measurements produced by Spranic. Each financial record ultimately rests on a physical asset whose condition can change between assessments.

An underinsurance example sharpens the problem: Spranger cites research covering 1,500 properties in which roughly 70% of buildings were underinsured. He explicitly leaves the cause unresolved. Old or insufficiently detailed property records are his proposed explanation, rather than a conclusion established by that study. The product question is therefore narrower and actionable: how can a building's observed state become easier to update and inspect?

“What would it take to make a building as queryable as a ledger?” A ledger can return an answer in milliseconds; a building question may require a site visit followed by a search through notes. Spranic's proposed input is an off-the-shelf sensor carried through the property. In the conference-hall example, Spranger describes walking through the room and having the scan a minute later. The useful relationship to inspect is between the captured space and its database representation: as the map forms, recognized assets are stored in a graph with semantic meaning.

Semantic meaning makes the record more useful than a collection of pixels. It represents what is in the space and allows information to be connected to those assets. The reason given for using a graph is flexibility: different properties can collect different information and connect it in ways suited to their assets. The talk explains that data-model choice without specifying the recognition models, graph schema or object-matching algorithm.

A repeat walk adds history. Both visits are retained and aligned along a time axis, whether they occur a quarter or a year apart. The record can then describe a space, its contents and its observed evolution. Its freshness still depends on someone walking and measuring the property again; the mechanism described is a sequence of observations rather than continuous monitoring.

The next requirement is a record that can support decisions. Spranger calls these “ledger questions” and separates three useful demands:

• What changed? Compare two walks of the same space object by object and against their surroundings.
• When was it observed? Preserve each observation's origin so a reported change can be traced to its capture. With separated visits, this establishes an observation history; it does not necessarily identify the instant the physical change occurred.
• Who needs to know? Make an observed difference available to the people who can respond to it.

Structural damage and expired fire extinguishers illustrate the proposed action layer. A finding with evidence could prompt corrective work or inform an insurer's assessment of property upkeep. These examples introduce intended uses of the record; the kitchen demonstration that follows tests the simpler, concrete case of recognizing objects and locating a change between visits.

The kitchen example begins with two captures, presented side by side. The system recognizes them as the same space and aligns time period A with time period B. That shared spatial reference is what makes the visits comparable: each capture becomes an observation of the same environment along a timeline.

What does querying a semantic map actually return? The demonstration uses two ordinary requests:

• Find a way to warm up. “Where can I heat myself?” points to the radiator in the map.
• Find something to drink from. A question about a glass or mug points to glasses and mugs in the kitchen.

The first request is especially instructive because it names a need rather than an object. The returned location connects that need to a recognized asset. In both cases, the map supplies a place to look, making the answer spatially useful.

The next query asks for the delta between the two kitchen passes: what changed in this environment? The reported result is specific—three chairs in the middle have been displaced. The answer also provides a reference back to the actual capture. The useful comparison is therefore both spatial and evidential: which objects moved, and which observation supports that answer?

This is the demonstrated meaning of spatial memory. The system keeps two observations, relates them to the same space, identifies an object-level difference and makes the supporting capture retrievable through a query. The example establishes that concrete capability; it does not establish detection accuracy across buildings or performance on safety-critical defects.

The closing section returns to the conference-hall scan and asks what other work could consume the same record. Proposed uses include repairs, maintenance, compliance checks, insurance and interior design. Robotics adds another possibility: a robot could use the representation as a digital twin of the space. These are downstream applications of the observed asset record, rather than separate demonstrations of completed integrations.

Insurance gives the proposal a concrete timeline. A walk at the start of a policy records the property's initial state. A second walk at renewal reveals differences and evidence of upkeep. An insurer could compare those observations with facts relevant to its policy and use them to price coverage more specifically. The same history might support a higher premium for poor upkeep or a more favorable assessment of a well-maintained property; the argument is for pricing based on observed conditions.

The business case is that insurance depends on facts someone can verify. Spranger closes with three external indicators:

• Construction classification. He cites one in three commercial buildings being recorded under the wrong construction class.
• Certification incentives. He describes hurricane-premium deductions in the 35–60% range for certified buildings, although the spoken account is unclear about the percentage's denominator and eligibility scope.
• Claims experience. He cites 73% of insurers reporting fewer hurricane-damage claims for checked and certified buildings.

These figures motivate better property evidence. They are not measured premium reductions or claims outcomes from Spranic.

Certification introduces a further requirement: observations must establish the particular compliance facts a program needs. Spranger proposes using AI to check such facts from the digital asset record and extending that approach to other qualifications inside a building. The recording demonstrates neither an insurer deployment nor a completed certification workflow, so the economic outcome remains a proposed application of the scanning and comparison mechanism.

The final ambition is a more transparent, property-specific market built on knowing the building's actual state. The moved chairs provide the small working example behind that ambition: preserve visits, compare the same space over time and return the observation behind a reported difference. Maintenance teams, insurers and other systems would then have a shared record from which to ask their own questions.