Commercial real estate · Local & private AI

Local and private AI for commercial real estate

Abstracting a lease is exactly what a language model is good at, and the documents are exactly what a firm has agreed not to distribute. Leases, LOIs, and ownership financials carry confidentiality obligations that a general-purpose API quietly conflicts with, which is why the most useful application in this vertical is also the one most firms have not deployed. Running the model on infrastructure the firm controls resolves the conflict rather than managing it.

10Self-hosted inference · Retrieval on private data
Local & private AI10
Self-hosted inferenceRetrieval on private dataDe-identification
Commercial real estateIllustrative

Local & private AI models

Self-hosted inference, fine-tuning, retrieval on private data, and de-identification.

Local & private AI · Commercial real estateOne pairing of the matrix
Covers
Brokerage & advisory · Owner-operators · Investment & syndication · Industrial, office & retail
Architecture
Unchanged from every other vertical — only the breaks differ
Engagement
A defined build, or run under management

What usually breaks

Before anything gets built.

The same failures recur across commercial real estate. Strygon maps the specific system before proposing anything, but these are the ones worth checking for.

The pipeline is somebody's file

Deals live in a producer's own spreadsheet with their own stage names. Nothing rolls up to the firm, comparison across producers is impossible, and a departure takes the pipeline out of the door with it.

Notice windows pass quietly

Expirations, renewal options, escalations, and estoppel deadlines sit inside lease PDFs. Nothing watches them, and a window missed by a week is a rent number that never corrects.

The quarterly package is a rebuild

Ownership and investor reporting is reconstructed every quarter from a rent roll, a bank export, and recollection. It arrives late, it is hard to defend line by line, and it consumes the people who should be working deals.

Recurring in commercial real estateObserved pattern · the specific system still gets mapped

The approach

What Strygon does about it.

Local & private AI models, aimed at those failures specifically rather than at a generic checklist.

01Local & private AI
Abstracts produced inside the firmLocal & private AI
01
Part of the scoped build

Abstracts produced inside the firm

Leases and amendments are read into structured abstracts on self-hosted models, so a confidential document is never handed to an outside processor to be understood.

02Local & private AI
Uncertain fields stay blankLocal & private AI
02
Part of the scoped build

Uncertain fields stay blank

Anything the model cannot read with confidence is flagged for a human rather than filled in, because a wrong escalation date is worse than a missing one.

03Local & private AI
Retrieval across the firm's own documentsLocal & private AI
03
Part of the scoped build

Retrieval across the firm's own documents

Questions are answered from the actual lease file with a citation back to the clause, instead of from a general model's impression of how leases usually read.

The approach for this pairingScoped in writing before work starts

What gets built

Concrete deliverables.

Everything in local & private ai models, applied to how commercial real estate actually runs.

Self-hosted / on-prem LLM and vision deployment
Fine-tuning, evaluation, and quantization pipelines
Retrieval (RAG) over private data
On-device and edge inference
De-identification and policy guardrails
In scopeScoped in writing before work starts

Also for this industry

Other parts of the same system.

Start

Start with what’s broken.

Send the situation in a paragraph. Strygon comes back with a read on what’s likely wrong and what it would take to fix, before anyone talks about price.

Most builds start withleads dying in an inbox., three half-finished pipelines., follow-up nobody owns., numbers that never agree., four vendors blaming each other.

What to send
A paragraph. What broke, and where it shows up.
What comes back
A read on what is likely wrong and what fixing it takes.
Price
The last conversation, not the first
What are you looking for?

Pick as many as apply.