Research

My research field is the management of knowledge about real-world things. The field includes entity resolution and record linkage, de-duplication, master data management and the integration of data across systems. Data provenance and lineage say where a fact came from, temporal data says when it was true, and a knowledge graph holds how things relate. I have worked on these problems for more than twenty years, in several information systems I have built. Most organizations have not solved them, and AI now offers easier ways to approach them than the tools I had before.

The same thing, described many times

Many organizations describe the same real-world things: a company, a product, a place, a person. Each of them keeps its own record under its own identifier, and no shared key exists. The records disagree and differ in age. A matching label does not prove a matching meaning: two systems can both have a field called “size” and measure different things.

Entity resolution is the work of deciding which records describe one thing. A shared name or a shared address is not enough for that decision, and a wrong merge destroys information that is hard to get back.

The same problem inside one organization

The problem also exists inside a single organization. The customer system, the billing system and the spreadsheet of a department each hold their own copy of the same customer, and the copies drift away from the record they came from. Every import adds duplicates, so de-duplication is a standing task.

A fact often sits in a system without its source and without its date. Nobody can say who entered it or whether it still holds. When the people who knew leave, the knowledge leaves with them.

What I work on

My answers have the same parts in every system. Each thing gets one identity that holds across records, systems and decades. Every fact is kept as a statement with its source, its author, the time it was observed and the time it was true. A correction adds a statement and overwrites nothing. Sources get a trust rank, so a conflict between two of them is decided by a rule that can be explained.

Records that describe the same thing are linked and keep their own data, which makes a wrong link reversible. Definitions, taxonomies and matching rules are versioned content that an editor can review, and they stay out of program code. Other systems read and write through one interface, under the rules a human editor follows.

What AI changes

For most of these twenty years the expensive part was manual: reading unstructured sources, comparing records, extracting facts and keeping definitions consistent. Language models do this work at a cost that a small team can pay. A model reads a document and proposes the facts in it. It compares two records and gives its evidence. An agent does curation work that nobody had the time for.

The limits are as clear. A model invents facts and merges what does not belong together. So a high match score never overrides a hard contradiction, every proposed fact names its source, a partial extraction stays marked as partial, and a human decides at the gates. The agent is a replaceable worker, and identity, rules and provenance stay in the store. My second field follows from this: how a company works when most of its staff are software agents.

How information moves through such a system

The diagram shows the path I build in every project. The pages below it take the parts one by one: the question, the model I arrived at, what is built and measured, and what is still open.