A $3.4B US real-estate investment firm
One honest answer box over a firm's entire portfolio
We built an internal intelligence platform for a real-estate firm managing about 3.4 billion dollars. One natural-language box answers questions about the portfolio and abstains rather than guess. Honesty is the product: every number is read from a real record.
Sector: Real estate, ~$3.4B AUM
Where this comes from
One production system, on the eval battery run at every release. It measures whether a refusal was correct, not whether the system refused often enough, so ask for the refusal rate beside it.
Where this comes from
Same system. A leak test only covers the tenant pairs someone wrote a test for, so this is a floor, not a proof of impossibility.
Where this comes from
Same system. Corpus size is evidence the approach holds at scale and says nothing about answer quality on its own.
Where this comes from
Same system. A count of specialist agents is an architecture fact, not a quality one.
The situation
- Thirteen years of portfolio data, mostly unstructured, that no one could query fast.
- Executives needed answers they could act on, so a wrong number was worse than no number.
- Every answer had to keep one property's data out of another's, with proof.
What we did
- 01A hub orchestrator classifies each question and dispatches the minimal set of specialist agents, then merges their work into one cited answer.
- 02Hybrid retrieval with a reranking stage, and a seven-stage verifier that byte-checks every claim against its source before it is served.
- 03A hard eval gate on every change: abstain precision, cross-tenant leakage, and golden wrong-values all have to clear a threshold or the change is reverted.
What held
- Live and in internal use, running on AWS Bedrock with a separate no-leakage island for legal.
- Real portfolio wired in: 33 properties, 9,521 units, 6,553 leases, all cited.
- The whole honesty battery stays green: zero hard-check violations, zero golden wrong-values, zero determinism violations.
Start a project
Name the workflow that eats your team's week
Name the workflow that eats your team's time. We will tell you honestly whether AI fits, and if it does, scope a first piece with a known price.