Private AI Search for 637,358 Public Records
Venkatesh Shivandi
The challenge
A public-sector property system held answers across 637,358 sensitive property and ownership records, but routine questions queued behind an analyst. Hosted AI was not acceptable because the data could not leave the client environment.
What I built
I built a private assistant that turns plain-English questions into executed database queries, then explains the returned results. It runs on client-owned hardware with no third-party API or network call.
Result
637,358 records searchable in plain English
Zero bytes sent to outside services
Routine questions no longer required an analyst queue
The system answers from real query results, so it cannot invent records that are not in the database.
Constraints
A local model is slower than hosted AI and is best for retrieval questions. Judgement-heavy work still needs an analyst.
Like this project
Posted Sep 2, 2026
A private AI assistant made 637,358 public-sector records searchable without data leaving the client’s hardware.
Likes
0
Views
0
