IronTinker

The abnormal record behind the incident, pointed out with evidence.

IronTinker is an AI support engineer for legacy business software: ERPs, point-of-sale systems, in-house apps, even decades-old core systems. It learns how the software really works from its data, then traces an incident back toward its cause.

Symptom"The supplier invoice match fails."
Unit prices of the items involved
ItemCreatedUnit price
A-102312016-03-024.20
A-102322016-03-023.95
A-102332016-03-0941,900.00
A-102342016-03-114.10

Out-of-pattern value: the price of A-10233 falls outside the column's range. Evidence:

SELECT ref, unit_price FROM items
WHERE ref IN (SELECT item_ref FROM invoice_lines
              WHERE invoice_id = 88412);
Illustration, fictional data.

Design commitments

Read-onlydesigned to never write to the target database
Evidenceevery conclusion must cite its queries and values
Masked datapersonal-looking columns masked by default before any AI call
By invitationask for access in two lines

The problem

Every support team running old business software knows the drill: a job fails with a cryptic error, and nobody knows why. Hours later, someone finds the culprit.

  • The daily sales of a till never reach head office: one malformed barcode blocks the export.
  • An invoice match fails with a cryptic error: one absurd price on a ten-year-old item.
  • Two documents that should be linked no longer are: a reference is missing.
The documentation never helps: the real behavior of the software lives in its data. In all three cases, the cause is one abnormal record upstream. IronTinker learns the software from its database, and later from its logs, docs, tickets and code.

Probe, the first module

Probe is designed to connect in read-only mode to a relational database. What it is designed to do:

  1. Explore

    It maps the tables, columns and keys, including the links nobody declared, and learns what normal data looks like, column by column: ranges, formats, valid EAN check digits.

  2. Trace an action

    Snapshot before, the user clicks in the app, snapshot after: Probe shows the changes it observed and explains them in plain English.

  3. Diagnose

    You describe the symptom in plain words; Probe walks to the rows involved and points to the value that breaks the pattern, with the exact queries and values as evidence.

  4. Learn

    Each trace and each diagnosis judged right enriches the map of the software.

Probe says what it knows and what it does not: an anomaly can be proven while its link to the incident is not, and it says so. A trace shows the changes observed between two states of the database, not a complete list of what the action wrote: other activity can mix in, and a row created then deleted in between stays invisible.

Safety comes first

  • Designed to never write to your database

    Three planned locks: a required read-only account, a read-only session, and a guard that rejects anything but reads. Probe is designed to check its own rights before it starts.

  • Evidence, or a hypothesis

    A conclusion must cite its queries and values. Without them, Probe must show it as a hypothesis, never as a fact.

  • The minimum goes to the AI

    Probe is designed to send the model only the schema, statistics and a few sample rows, and to mask columns that look like personal data by default.

  • Fictional data for demos

    The public demo is planned on a demo ERP with fictional data: corrupt the price of an old item, watch a reconciliation fail, and see Probe look for the cause.

Probe is designed to read the history of your software in its data.

Access is by invitation. Write to us in two lines: the kind of software you maintain and the database behind it.

Request an invitation