Skip to content
Ilayer

We put AI into production

Building AI is the easy part.

The hard part is your systems. We build inside your own code, on your real data, and we stay on call for what runs. First working system in four to six weeks.

What we take on

We ship the tests with the system, and every number on this site maps to the run that produced it. How it is measured

Three separate pieces of work, each measured

1.000Right every time it said it did not knowWhen the answer was not in your records, it said so instead of guessing. It scored 1.000 on that: every time.This counts only the times it held back. It cannot tell you how often it answered when it should have held back.
60%Cut off a production cloud billOne audit on one cloud estate: $5,506 a month down to $2,186, with no new capacity bought.One estate, one change window. That is what that account was paying, not a rate we quote.
29M+It holds up at real sizeMore than 1.3 million documents, broken into more than 29 million searchable pieces.Size shows the method holds up at volume. It says nothing about whether any one answer is right.
7-gateFindings filtered before a person sees oneSeven checks every possible finding has to clear before a person sees it, on the security research system we built for ourselves.Our own system, not a client's. A check can filter out a finding; it cannot prove one is exploitable.

Each of these comes from a different piece of work: a live system we still run, one cloud estate we audited, and our own security research. None of them is an average across clients. One of them went down. We re-measured the source and published the correction.

The work

Six parts of one system. Only checked work reaches production

Everything goes through the same checks. Work that fails a check stops there.
  • AI agentsSoftware that does a whole job, not one step of it.
  • Knowledge and answersAnswers from your own documents, with the source attached.
  • AutomationThe same steps every time, done without a person.
  • Testing and reliabilityChecks that catch a wrong answer before you do.
  • AI securityAI that tries to break code before an attacker does.
  • InfrastructureThe cloud it all runs on.

That is the work. Four ways to buy it, each priced before we start.See the four

How it works

It looks the answer up instead of inventing one

Ask a general model what it does not know and it answers anyway. This one looks it up first, and stops when there is nothing there.

Your sourcesThe gateCited, or nothing
  1. 01

    Reads your sources first

    Your documents, databases and systems, not what a model picked up in training.

  2. 02

    Answers only from what it found

    Every line links to the document behind it, or it says it does not have one.

  3. 03

    Checked before it reaches you

    A verifier blocks any claim it cannot trace to a source.

Honest about the limits. Grounding does not make mistakes impossible. It makes them visible instead of silent.

What an answer looks like

Two outputs. One is a refusal, and that is the point

Illustrations of the shape, not client screenshots. The figures are withheld because they would be a client's. The shape is the same whatever the records are.

Grounded

What did we commit to on renewal terms last quarter?

A figure for every agreement. Each one carries the record it came from. The ones outside the usual range are named, with the reason.

Every figure links to the row it came from.

Refused

What will those terms be next quarter?

I do not have that. Nothing in the records is a forecast.

No source, so no number.

The refusal is the harder problem, and the one most systems skip.

A dark cylindrical body with one thick pipe entering from the left and five smaller outlets fanning out to the right, each outlet glowing violet at its mouth.
The stack underneath is already yours.

Your stack

It runs on the stack you already have.

We build on the models, data and cloud you already pay for. Nothing gets ripped out.

How an engagement runs
  • ClaudeModel
  • AWS BedrockRuntime
  • OpenAI embeddingsSearch
  • Voyage rerankerSearch
  • AWSCloud
  • PostgreSQLData
  • pgvectorData
  • FastAPIFramework

No endorsement implied.

Selected work

Real systems, real numbers

Anonymized by default.

All work
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.
A $3.4B US real-estate investment firmThirteen years of records. One box that answersThirteen 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.1.000Correct every time it held back0No client's data in another's answersRead the case
Seven thin machined gates in a row on a dark optical rail, receding to the right. A violet beam enters the nearest aperture and passes through all seven, ending in a bright point past the last gate.
Ilayer researchAI that reviews code the way an adversary wouldA tool that cries wolf costs more time than it saves. We built a system of several AI agents. It audits smart contracts, the code that moves money on a blockchain. It attacks its own findings, and shows a person only what survives. It has found real vulnerabilities on live protocols.7-gateEvery finding filtered before a human sees itLiveReal bugs on live protocolsRead the case

Straight terms

What you can count on

We tell you when AI is the wrong tool

You hear it before you spend. Where this shape does not fit is written down, not implied.

You know the price before we start

One price for the build, agreed before it starts, then a flat monthly rate.

You own everything we build

Your code, your license, runnable without us. The test is whether you can end the monthly rate and keep the system running.

Questions

What buyers ask first

What does Ilayer do?

We take one workflow that costs your team real hours and get it running inside your own systems. Then we keep it running. Same engineer, same tests, and a gate that blocks a change that makes the system worse.

How long does a first project take?

A fixed-scope first piece runs 4 to 6 weeks, kickoff to something working on your real data.

What does an engagement cost?

Fixed-scope build, priced before we start, then a flat monthly rate. No open-ended retainers.

How do you stop AI from making things up?

It answers from your documents rather than from memory, shows the source, and says so when the answer is not there. On one live system, every time it said it did not know, it was right to: a score of 1.000.

Five glass rings in a row, receding and shrinking to the right, one violet beam passing through every center, flaring brighter at each ring and ending in a small point of light.
Every claim on this page has passed the same checks the product does, or it is not on the page.

Start a project

Name the workflow. We will tell you if AI fits

One paragraph is enough. You get an honest read on whether this is an AI problem, and what a first fixed-scope piece would cost. If you do not need us, that is the answer you get.