Intelligence · Smart Bricks models
The engine behind every number.
7 in 10 agent tasks run on our proprietary models, tailored for real estate and hosted on our own GPUs. The rest are routed, task by task, across 30+ frontier AI models.
Request a DemoThis runSmart Bricks models 7 of 10Frontier 2Engine 1
A sample run on our sample deal. Select a task to see why it went where it did.
What the models are
Our models have read the asset class.
Smart Bricks proprietary models
- Tailored for
- Real estate
- Hosting
- Self-hosted, on our own GPUs
- Developed by
- The Smart Bricks team
- Agent tasks
- 7 in 10
A frontier model
Has read the internet
- OpenAI
- Anthropic
Handles language
A Smart Bricks model
Has read the asset class
- Lease documents
- Rent rolls
- Title records
- Completed transactions
Handles real estate
Training data
Fine-tuned on lease documents, rent rolls, title records and completed transactions.
Each transaction enriches our data, and that data is what we fine-tune the next model generation on. Client data and decisions join it only with that client’s consent.
- Lease documentsLease-level economicsIncluded
- Rent rollsTenants, rents and expiriesIncluded
- Title recordsResolved ownership chainsIncluded
- Completed transactionsConfirmed pricesIncluded
- Off-market supplyWhat no feed sellsIncluded
- Client data and decisionsWhat each client decides on the platformOnly with that client’s consent
Routing
No dependency on a single lab.
If a model provider changes price or terms, your work does not stop: each task reroutes to the next best model, and your workflows, memory and audit trail stay put.
- Smart Bricks modelSelf-hosted
- OpenAIFrontier, routed per task
- AnthropicFrontier, routed per task
- GoogleFrontier, routed per task
- Open modelsLlamaMistralQwenDeepSeek
- and more, routed per task
Every route is chosen on
- Quality
- Cost
- Latency
- Fallback
- Compliance
- Evals
- Provider degradesRerouted to the next best model
- Provider repricesRerouted on cost
- Provider restricts accessRerouted on compliance
Financial Engines
The model proposes, the engine computes.
Cash flows, debt schedules, waterfalls and returns run in coded financial engines, never in a language model, so the same inputs always give the same number.
- cash flows
- debt schedules
- waterfalls
- returns
- Levered IRR
- 14.8%
- Equity multiple
- 1.9x
- Downside IRR · ERV -10%, exit +75bps
- 7.1%
- Smart Bricks modelAI where intelligence matters.
- Deterministic engineDeterministic engines where precision matters.
Model validation
Every model is tested on transactions it has never seen.
Error is measured against confirmed transaction prices, not list prices or automated estimates, market by market, and the models stay under watch in production.
- 01
- 02
- 03
- 04
- 05
- 06
Train Test
- Split
- Temporal. The test set is the most recent transactions, never exposed during training.
- Walk-forward
- Validation across 6-month rolling windows, to prevent lookahead bias.
- Error
- Median absolute error against confirmed transaction prices, not list prices or automated estimates.
- Markets
- Separate sub-models per market, unified by a meta-learner for cross-market signal.
- In production
- Tracked against closed transactions since January 2024. Drift detection triggers retraining.
- Median absolute error
- 4.5%
- Hit rate
- 94%
Across New York, London and Dubai, on transactions the models never saw in training. Hit rate is the share of valuations that land inside the model’s confidence band.
| Measure | Generic LLM | Smart Bricks Models | Difference |
|---|---|---|---|
| SpeedFinancial analysis, 76 files | 255s | 80s | 3.2x faster |
| TokensSame run, total consumed | 69k | 48k | 1.4x fewer |
| AccuracyComplex modelling tasks done right | 43% | 95% | 2.2x more accurate |
Smart Bricks internal benchmark vs general-purpose agents: our own model completes real estate workflows faster, at lower cost and with materially higher accuracy.
In a real deal
It all lands as finished work a committee can act on.
The sample deal as it reaches the investment committee: what came back, and which kind of model did the work.
Smart Bricks Investment Score
87 / 100
Evidence confidence 95%
Entry yield of 5.4% stabilising to 6.2%, a 14.8% five-year levered IRR and a 1.9x equity multiple on a £385m acquisition. The downside case (ERV -10%, exit yield +75bps) holds a 7.1% IRR.
- WAULT6.3 yearsSmart Bricks model
- Top 5 tenants61% of NOISmart Bricks model
- Entry yield5.4%Smart Bricks model
- Stabilised yield6.2%Smart Bricks model
- Levered IRR14.8%Deterministic engine
- Downside IRR7.1%Deterministic engine
- Investment recommendation
- Underwriting and scenarios
- Maximum supportable price
- Risks and exceptions
- Client Excel model
- IC materials
IC review
A frontier model has read the internet. Ours have read the asset class.