Expert finder

Halfway through a deep-dive, a team usually knows which questions the literature will not answer. Somebody has to talk to a person. The obvious names are the ones already quoted in every review. The useful name is often a technician, a regulator, or a postdoc who published once and moved on.

An expert finder returns a candidate list. Each row carries a name and a current affiliation, with the date it was checked. It also carries the specific work that makes the person relevant, and a link to it. A row without the underlying artifact is a suggestion, and suggestions are what the team already had before it ran anything.

The list is not a judgment about expertise. It is a set of leads ordered by how much evidence exists that this person has done the exact thing the deep-dive is stuck on. Somebody still reads the work, decides whether the person is worth an hour, and takes responsibility for the invitation.

This is the cheapest step in the stage and it stays cheap only if the team resists turning it into a ranking exercise. Ten names with real artifacts beat a hundred scored on prominence.

Start from the gap, not from the field. "Find experts in antimicrobial resistance" returns the same twenty people for any question in that field. Most of them cannot help with the uncertainty in front of the team.

A usable brief holds three things. The sub-question, written as a sentence somebody could answer in a conversation. The kind of person who would plausibly know, meaning a bench researcher, a field epidemiologist, a farm veterinarian, an inspector, or a data custodian. And the disqualifiers: an active commercial interest, a role inside an organization the review is assessing, or a language and time zone the team cannot work with.

Add the sources the search may read. Journals, conference programs, citation graphs, patents, standards committee rosters, and work profiles each behave differently. Each fails differently too. A search over journals alone misses people in practice, and those are often the ones the deep dive needs.

When the sub-question is missing, return nothing. A name list generated against a field label wastes the outreach budget on people who will politely explain that the question is outside their work. The same instinct drives targeted sweeps of practitioner reports and registries. The interesting evidence often comes from people who never wrote a paper about it.

Prominence and relevance are different measures, and most tools default to prominence because it is easier to compute.

Rank on four signals, and keep them in separate columns. Topical proximity, meaning how close the person's own output sits to the sub-question, measured against the actual text of their work rather than a keyword on their profile. Recency, meaning whether the relevant work happened inside a window where the field has not moved on. Role fit, meaning whether their position lets them speak to the question at all. Reachability, meaning whether a current contact route exists.

Semantic retrieval over the right corpus does the topical part. Semantic search over the appropriate research corpus is the same machinery, pointed at people rather than papers. Feed it the deep-dive's open question in full sentences. A three-word query returns celebrities.

Name disambiguation is the unglamorous half of the job. Common surnames, initials, transliterations, changed names, and shared institutional pages will merge two researchers or split one into three. Run vector similarity to merge near-duplicates and aliases over the candidate rows before anyone reads the list. Keep the merge decisions visible, so a wrong merge can be undone.

Show the evidence next to the score. If a person ranks third, the row should say why. Two papers on the exact assay. A talk at last year's meeting. A current post at a national reference laboratory. A rank with no artifact behind it cannot be checked and will not survive a colleague asking who this is.

The following is an invented scenario. The names, counts, and affiliations are illustrative.

A research unit is weighing a change to a national drug-resistance watch program in poultry. Today it samples at slaughterhouses. The question is whether to sample on farms as well. The published work covers the disease science well. It says almost nothing about whether farm sampling can be run at the scale the design needs.

The brief splits into three smaller questions. Can farm sampling cover small producers fairly. Do the labs have room for the extra samples. And how did farmers react where a similar program was tried before.

The search runs over journals, two conference programs, a standards roster, and public work profiles. It returns 38 names. Merging cuts six pairs. Two were spellings of one surname. One was a researcher listed under an old employer.

Reading the evidence columns, the list divides. Eleven candidates are cited heavily for mechanism work on resistance genes and cannot speak to sampling logistics. Nine are farm veterinarians and regional inspectors whose relevant artifact is a technical report rather than a paper. Four sit on the standards committee that wrote the sampling guidance. Two ran a farm-level pilot in another country and published only a conference abstract about it.

The unit invites six people: the two pilot leads, two inspectors, one laboratory director, and one committee member. The heavily cited mechanism researchers, who would have topped a citation-weighted ranking, are not contacted at all, because none of the three sub-questions is about mechanism.

Every failure in this step produces a plausible name, which is what makes them expensive.

Stale affiliations are the most common. People retire, move, take industry roles, or die, and profiles lag by years. Record the date each affiliation was checked, and treat anything unverified in the last year as unconfirmed on the row rather than as fact.

Fabricated evidence is the dangerous one. A model asked to justify a name will produce a paper title, a coauthor list, and a plausible venue for work that does not exist. Every row needs an artifact somebody can open. If the link does not resolve, drop the row instead of asking for a replacement citation.

Prominence bias narrows the room. Citation counts, conference invitations, and media presence all reward the same people, and a shortlist built from them reproduces the consensus the deep-dive is trying to test. Deliberately read the lower-ranked rows for practitioners and regional specialists, the way mapping what peers are scoping toward forces attention outside the obvious set.

Conflicts of interest do not show up in a citation graph. Advisory board seats, funding sources, expert witness work, and equity are disclosed inconsistently or not at all. Check them by hand before an invitation goes out, and record what was checked.

The researcher owns the contact decision and everything after it. Automated outreach from a list nobody read burns a team's standing in a small field. Small fields are exactly where this method is most useful. Write the invitation as a person, name the sub-question, and say how the conversation will be used.

Once someone agrees, the finder's job is finished. The question set, the recording, and the synthesis belong to tailored question sets and post-interview synthesis. That step takes the sub-question and the artifact from this row as input. Handing over a name with no context produces an interview that repeats the literature review out loud.