Public Sector · Private AI · 4 weeks

They could ask 637,000 records a question in plain English. Nothing left the building.

The obvious tool was off the table, because sending government property and owner records to somebody else’s server was never going to be approved.

Zero bytes

Sent to any outside service — the model ran on hardware the client owned

Their property system could answer almost anything, if you knew how to ask — the right combination of five or six filters. In practice every question queued behind one analyst, so things that would take thirty seconds to ask took days to answer.

We built an assistant that takes the question in ordinary English, turns it into a real database query, runs it, and then explains the result. The model runs on their own machine. There is no API key and no network call, so no data leaves.

The part that mattered for approval: because every answer comes from an executed query rather than the model’s recollection, it cannot invent a property that is not in the records. Ask about something that does not exist and it says so.

A state-chartered land bank authority, US Northeast

What it came down to

Answers come from real queries, so the system cannot invent a record
Runs on client hardware — no third-party service in the loop
Staff stopped queuing behind an analyst for routine questions

Records made searchable: 637,358
Records made searchable
Sent to any outside service: Zero bytes
Sent to any outside service
From no system to a working one: 4 weeks
From no system to a working one

What this doesn’t tell you

A local model is slower than a hosted one and needs a decent machine. For questions that need judgement rather than retrieval, an analyst is still better.

Clients are described by sector, scale and region only. No names, products or identifiers appear here.

Next step

Have a version of this problem?

Tell us what you are trying to build. You will get an honest read on whether it is worth building at all.

What happens next

You pick the workflow. We ask what already exists, who touches it, and where it stalls.
We reply within one business day with a rough shape and a rough timeline — usually smaller than people expect.
If it makes sense, the next step is a working system in two to six weeks, tested by your own staff.
You keep your existing tools. Private deployment is available where data cannot leave.
And if it is not worth building, we say so instead of selling you one.