Skip to content
Ilayer

A $3.4B US real-estate investment firm

Thirteen years of records. One box that answers

Thirteen years of records nobody could query fast. We built one box that answers across all of it, cites every number, and refuses when the data is not there.

A wall of identical dark filing drawers photographed straight on. One drawer left of centre is pulled fully open, the tabbed index cards inside it lit violet, the glow spilling down the drawer fronts below.

The situation

  • Executives act on these answers: a wrong number is worse than none.
  • One client's data had to stay out of another client's answers, with proof.

What we built

  1. One coordinator sends each question to one of 5+ specialist agents, each holding a different part of the data.
  2. Every claim is checked against its source before it is served.
A large violet glass sphere held in a dark gimbal ring on a knurled base. One beam of light enters from the left, fans into refracted rays inside the sphere, and leaves to the right as a single focused line.
One question in, refracted across thirteen years of records, one cited answer out. Illustration, not a screenshot.

Measured

What this system holds

1.000Correct every time it held backIt scored 1.000 for this: when the data was not there, it said so instead of guessing.One production system, at every release. It does not measure how often it should have refused.
0No client's data in another's answersOne client's records never turned up in another client's answers. Checked in both of the places the system stores its search index.We only tested the client pairs somebody wrote a test for. That is a floor, not proof.
29M+Holds at real data volumeMore than 1.3 million documents, broken into more than 29 million searchable pieces, every one of them citable.Scale shows the method holds, not that an answer is right.
A dark oscilloscope in a machined case with two round knobs and a vent below the screen. One violet trace wanders across the screen from edge to edge.
One reference line, and the system measured against it at every release.

What held

  • Live and in daily use. The honesty checks still pass at every release.

The stack

  • Claude on AWS Bedrock
  • OpenAI embeddings
  • Voyage reranker
  • PostgreSQL + pgvector
  • FastAPI

Built with

Evals are tests for a large language model's answers. RAG is short for retrieval-augmented generation.

AI agents and answersTesting and reliabilityEmbedded AI engineeringAll work

Start a project

Ask what this would look like on your records

The system above was one workflow, scoped before it started. Name yours and you get a straight read on whether your data supports it.