A read-only layer over grown IBM i (AS/400) and HCL Domino estates. It reads the catalogue and the source, builds a searchable model of the system, and answers questions about it — each answer carrying the exact place it came from.
How much of your business logic is still readable?A one-off inventory analysis from your own export — one number instead of an opinion. The ordering page is in German; write to me and we will do it in English.Companies rarely replace IBM i or Domino because the platform stopped working. They replace it because nobody knows what the applications do any more: the colleague who built them has retired, part of the business logic exists only as a compiled object, and the documentation describes a state from 2009.
The decision to rebuild then gets made out of ignorance about the estate rather than for technical reasons. That loss of knowledge is what this product addresses.
ORDSTS is in VERTRIEB.ORDHDR and is written by program ORD410, statement 47.The source reference is not decoration. An answer about a core system that cannot be looked up is worthless when it matters — so before answering, the system checks that every statement is covered by its source, and discards what is not.
The property worth understanding before anything else: it can return nothing. „No provable answer in the indexed estate“ is a valid result. A tool that silently guesses about a core banking system is worse than no tool.
The first live run found three genuine defects that were green against the mock. A mock built from the same assumption as the code confirms the assumption; it does not test it.
From 16 public Notes applications with timestamps back to 1997 — real grown legacy, not laboratory data.
That is how little was present as source in that corpus. The agent detects this and says so, instead of pretending it read everything.
The language model and the search run on your premises. A target outside your own network is refused — checked against the resolved address, not a name.
Process mining needs three things per event: a case ID, an activity, and a timestamp. For SAP there are templates. For an RPG application written in the 1990s there is none — and the developer who knew has retired.
The agent proposes the three roles from what the catalogue gives up: case IDs from declared keys, activities from columns whose measured value cardinality is low, and timestamps including the IBM i idioms no data type reveals — 7-digit CYYMMDD, split xxxDAT/xxxTIM pairs. Each proposal carries its evidence and a confidence value, and it emits the Activity and Case Table specification plus the SELECT for an existing JDBC extractor.
Since September 2026 there is, separately, a write path for IBM i, and it belongs to a human: it is off in a new installation, runs over a second connection with its own user profile, requires a second person’s approval, and a change request carries table, key and field values — never SQL. What a human approves is therefore readable, and what runs is derived from it. It refuses to touch a file that is not journalled, because without a journal there is no rollback on IBM i.
An installation in your own house, air-gapped if you want it. Two cores and 2 GB of memory are enough for the service itself; an estate with 10,000 design elements occupies around 1 GB in the database. Only the language model is compute-heavy, and it may run on an AI server you already have. A 7 B model is sufficient — measured against a 14 B on the same corpus, the smaller one answered more questions correctly and caught every trap question.
The product is in active development, accepted against a real IBM i 7.5, and measured against a public corpus of real Notes applications. There is no customer reference yet — a sentence one is tempted to leave out, and leaving it out is what makes a first meeting go badly.
If you look after a house where the replacement question is on the table, or if you implement process mining and keep running into an AS/400 that nobody can document: I will gladly show the tool against one of your own applications. alexander.erb@agent-hub.app