The evidence layer for IBM Z

Ask your mainframe
anything.

Sessions, runbooks and COBOL become one retrievable memory. Every answer arrives with its sources attached.

Book a call Or email hello@getayori.com
Illustrative example of the Ayori interface. The estate, identifiers and counts shown are fictional.
Payments estate  /  Askgraph v4 · 12,481 facts
Corpus
Sessions41
Runbooks128
COBOL / JCL3,902
Graph facts12,481
Recent
Who owns the CARDHOLD retry?Batch window for NSB.SETTLE
What breaks if AUHOLD1 is unavailable at settlement?
Answer

NSB.SETTLE skips to the recovery step and flags the batch for manual review.1 It does not retry automatically: the retry flag is owned by CARDHOLD, not the settlement job.2,3

NSB.SETTLEUSESAUHOLD1READSNSB.CARDHOLD
Sources
1session_0421#c3Settlement walkthrough · 14:02
2runbook_settle.pdf · §3Recovery procedure
3NSB.CARDHOLD.cbl · L212Retry flag
Retrieval trace
0.89session_0421#c3
0.84runbook_settle#c7
0.79CARDHOLD.cbl#c2
0.31runbook_batch#c1
chunks kept3
graph hops2
prompt2,041 tok
All claims mapped to retrieved evidence.

The people who know how it works are leaving.

Knowledge lives in people

The engineer who knows why the settlement job retries twice is one retirement away from gone.

Documentation drifts

Runbooks describe the system as it was designed. Production has moved on without them.

Search doesn’t help

Grep finds the string. It won’t tell you what breaks three jobs downstream.

How it works

Four steps from tribal knowledge to a cited answer.

01

Capture

Expert sessions, runbooks, specs and code go in. Chunked, cleaned, embedded.

02

Connect

Relationships between programs, jobs and datasets become an explicit graph.

03

Retrieve

Top-k search, then graph hops. Only the evidence the question actually needs.

04

Answer

With its chunks and graph facts attached, or an honest “not in scope.”

Measured

Retrieve the evidence, not the estate.

Full-context baseline
126,952
tokens sent per query
Ayori, mean per query
~2,029
1,536 min · 2,660 max
98%
less context per answer

Both figures are real Bedrock input-token counts on a synthetic MVP corpus. A reduction in context sent, not a cost claim.

Sourced

Nothing to take on faith.

Each answer carries the chunks and graph facts it was built from. When the evidence isn’t there, Ayori says so instead of guessing.

ayori
Does CARDHOLD write to the audit file on a failed hold?
I don’t have evidence for that. The corpus covers CARDHOLD’s retry path but no captured session or runbook describes its audit-file behaviour on failure.
0 chunks above threshold

Deployed per client

Runs against your estate, not a shared multi-tenant service.

Every answer sourced

Chunks and graph facts travel with the answer, fully auditable.

Refuses when unsure

No evidence, no answer. Ayori tells you what it doesn’t have.

Questions we get.

Does Ayori need our source code?

Only what you point it at: sessions, runbooks, specs and the COBOL/JCL in scope. Nothing leaves the deployment.

Where does it run?

Packaged and deployed per client, against your estate. Not a shared service.

What happens when it doesn’t know?

It refuses. Answers are constructed from retrieved evidence only, so a gap in the corpus shows up as a gap in the answer, not a guess.

Are the numbers on this page from production?

No. They come from a synthetic MVP corpus: fictional test data measured by a reproducible harness. Cost figures are calculated token-level Bedrock costs, subject to AWS invoice verification.

What do you need from us to start?

A scoped estate, a few expert sessions to capture, and someone who can say which systems matter most.

Give your engineers
the memory they’re losing.

Book a call