Dagnul, Inc., Boston
Adverse-event intake for Drug Safety from streams other teams run
The work
Dagnul builds intake software for the Drug Safety teams of drug companies, starting with NDA, ANDA and BLA holders in the United States.
A serious, unexpected case has fifteen days from the day anyone at the company, or working for it, holds the four minimum criteria, weekends included, and many cases start in a medical-information reply, a call-center contractor's note, a brand-page comment or an agency's AE form, run by teams the head of pharmacovigilance cannot instruct.
Dagnul's queue takes a read-only copy of each stream the head of pharmacovigilance switches on and holds every possible case on the day it arrives. Each item shows the four criteria, a proposed seriousness and a proposed day zero beside their sentences; the owning team is copied, a safety scientist confirms or corrects, and the dated trail is kept for ten years.
Reading a patient's sentence for a case is work that keywords get wrong in both directions, so Dagnul trains a model for it.
Keywords flag every package-insert request and miss a hospital stay inside a dosing question. Our open-weight multilingual encoder, fine-tuned in English and Spanish, marks the criteria, the seriousness and the sentence that starts the clock; rules take the timestamps.
Dagnul's model may propose and route, and it may not close an item, even one it calls not a case.
Items below the line the scientists set for each stream go to a daily list a person confirms. Every correction is stored as one labeled pair, names replaced by placeholders, which belongs to that customer and retrains only its model, monthly, under the change control plan the head of pharmacovigilance signs.
Dagnul's GPU load comes in bursts, because reading one item is light work and three jobs are not. A new customer's backfill reads a quarter of logs at once, each monthly retraining reruns a held-out test set, and a brand campaign can fill a page with comments in a weekend. Patient text in a safety system should not reach third-party model APIs, so training and inference run on cloud GPU capacity we control in N. Virginia.
- January 2022Dagnul, Inc. was founded in Boston in January 2022 to count the days a case spends outside Drug Safety, at first with rules over intake logs and case lists.
- 2023The rules missed the sentences that mattered, so in 2023 contract pharmacists began writing and labeling our corpus, and in 2024 we trained the first model on it.
- 2026The service opened in 2026.
- March 2027Dagnul's hardware plan has two dates. In March 2027 retraining leaves hourly rental for GPUs booked for twelve months in N. Virginia, and in the second half of 2027 the model reads Portuguese, which our contract pharmacists began labeling this year. Text in any other language keeps reaching a person unread.
The people
Evelien Daems
CEO
Was QPPV and head of drug safety at a Flemish consumer-health company and signed CAPAs for late cases whose first days had passed in inboxes run by teams that did not report to her.
Frederik Leysen
CDO
Did his doctorate in Leuven on finding medication events in Dutch and French clinical notes, and owns Dagnul's corpus, its held-out test sets and every model version a customer signs.
Femke Verlinden
CPO
Ran a medical-information service for a generics group for seven years, answering patients' dosing questions under an SOP that gave her team one business day to pass a reaction on.
Dagnul, Inc., 451 D Street, Suite 702, Boston, MA 02210, United States
Get the count