KUnnect is a prototype intended to help us explore how to develop and deploy AI-based
applications at KU, particularly with the use of open models. Content is generated by AI and
may contain inaccuracies.
This application was developed by the Center for Health Data Science (HeaDS), on behalf of the
Deanery of the Faculty of Health and Medical Sciences. It has been approved for release by
KU-IT.
KUnnect runs on a single-GPU server in Denmark under a 100% renewable energy
policy. Building it cost an estimated 400 kg CO₂e, and a live query that
requires ~1 minute of GPU adds about 2 g each time. Estimated with
Green Algorithms
(Lannelongue et al., 2021).
Questions, corrections and ideas are welcome through
Contact.
KUnnect turns public publication records into meaning-based profiles, then matches
researchers working on similar ground. Here is the path from raw data to a recommendation.
Collect
KUnnect is built on publicly available researcher profiles from CURIS, the University of
Copenhagen's official research information system. The dataset covers approximately
7,000 researchers whose profiles were publicly indexed in CURIS as of August 2026
and who had at least three publications recorded since 2021. A profile also needs at least
one of those publications to have a readable abstract, so a small number of otherwise
eligible researchers are missing. Only publicly accessible pages were used; no login or
restricted database was involved.
Link
Each researcher is tracked through a unique profile identifier, so author names are never
guessed or confused between people who share a similar name. Their publications are linked
back to that single identity before any analysis begins.
Read
KUnnect uses a technique called semantic embedding to capture the meaning of each
researcher's work, going beyond titles to the content of the research itself. A language
model called Nomic AI reads each publication abstract and converts it into a list of 768
numbers encoding the meaning of the text. Trained on vast scientific literature, the model
knows that "myocardial infarction" and "heart attack" refer to the same concept, and that a
paper on Arctic ice cores is closer in meaning to one on permafrost than to one on corporate
tax law. Each researcher's publication fingerprints are then averaged into a single profile
representing the overall territory of their recent work. Only publications from 2021 onwards
are used, so the profile reflects what a researcher works on now rather than their full
career.
Compare
When you search for a researcher, KUnnect first identifies their existing co-authors from
papers published since 2021. KUnnect then searches the full pool of roughly 7,000 researchers for
those whose topic profile is closest to the queried researcher's, surfacing people covering
similar scientific ground who have not recently collaborated. Researchers who collaborated
with them prior to 2021 are not excluded and may appear as potential matches.
Co-authorship counts no matter how many authors a paper had, so a large consortium paper
can rule out someone a researcher has never worked with directly.
Recommend
A language model then reads papers from both sides and writes a short explanation of the
potential overlap — the recommendation text you see in your results. On a name search it is
shown up to 20 of that researcher's most recent papers and up to 10 from each candidate.
Name and theme searches use Qwen3.8 27B; the SUND collaborates grant tool, the
Describe research box and the Upload a PDF analysis use Gemma 4 31B. A smaller
model, GPT-OSS 20B, reads each submission first and declines the ones that are not
research descriptions at all. All are open-weight models running on our own server, and
the two that write recommendations were chosen by benchmarking candidates against each
other on real searches, scored for whether every person and paper they cite is one they
were actually shown.
Public data only
KUnnect is built exclusively on publicly available researcher profiles from CURIS, the University
of Copenhagen's official research information system. No restricted databases, logins, or
private records were accessed.
No permanent records with personal information
We do not store what you search for or submit. Uploads are processed in memory and never
written to disk; results are deleted as soon as you retrieve them, or within an hour if you
don't. We keep no IP addresses and set no tracking cookies.
To know how much the tool is used, we count visits and analyses. Each is filed under a code
derived from your network address and browser using a secret key that is destroyed every
night, so yesterday's counts can no longer be traced to anyone — not even by us. We record
which feature was used and when, never the name, theme, text or file you sent.
Descriptions submitted to SUND collaborates are screened by a small language model before the
analysis runs. Whether it accepted or declined, and the one-line reason it gave, are kept
alongside that count so we can tell how well the screening works; your description itself is
not. Server logs, which contain no search or upload content, are deleted after 14 days.
Contact form submissions are kept for six months so we can reply, then deleted
automatically.
Danish infrastructure
All computation runs on a server hosted in Denmark by
Scannet, under a 100%
renewable energy usage policy. We are considering options of hosting it in servers managed
by KU-IT in the future.
Local large language models
The language models used for embedding and generating recommendations run locally on that server.
No data is sent to external AI providers such as OpenAI or Anthropic. Recommendations are
written by Qwen3.8 27B for the main search function, and by Gemma 4 31B on the SUND
collaborates page and for the Describe research and Upload a PDF analyses. GPT-OSS 20B
screens submissions before they are processed. Publications are embedded with Nomic Embed
Text v1.5.