The work behind the answer.
Getting to a real answer takes more than generating something quickly.
Better, and more efficient.
It takes retrieval, reasoning, orchestration, compute, and systems that know how to recover when something goes wrong. Our research focuses on making that process better and more efficient.
- Focus
- Models · efficiency · agents
- Latest release
- BEM-2
- Stage
- Internal research
- Evidence
- Held-out, seeded, falsifiable
Our models.
Models we build and train ourselves, each one for a product.
Latest release
BEM-2
Our new family of biomedical models, and the successor to BEM-1.
Watching cells, not just reading about them.
Some questions live in the data rather than the papers. Kinema learns how cells move, meet and divide from time-lapse imaging, from far fewer examples than a conventional video model needs, so an analysis can reason about how a living system behaves over time.
- Area
- Cell dynamics
- Stage
- Research model
- For
- Essarion Science
A model that keeps learning.
Knowledge doesn’t stand still, and a reference shouldn’t either. DAI keeps learning after it ships: it updates its own weights as new material arrives and chooses what is worth learning next.
- Area
- Continual learning
- Stage
- In development
- For
- Sourcipedia
Screening sources at speed.
Good research starts with knowing which papers and sources are worth reading. Small, fast models screen them first, so the deeper reading goes where it counts.
- Area
- Paper and source screening
- Stage
- In development
- For
- ResearchAnything
BEM-1
Our first family of biomedical models, now succeeded by BEM-2. Released August 12, 2026. Read about BEM-1.
BEM-1 Examine
EncoderJudges whether biomedical text says more than the evidence behind it supports.
BEM-1 Examine Mini
EncoderA compact version of Examine for ordinary hardware. It handles the easy majority and passes harder cases up.
BEM-1 Compose
DecoderWrites in the register of a chosen kind of study, from case report to meta-analysis.
BEM-1 Critique
DecoderRaises the methodological questions a careful reviewer would ask of a claim or an abstract first.
How we measure our work.
A result without its measurement conditions is not a result. These are the rules we hold ourselves to.
Test on data the system has not seen
We report results on held-out evaluation windows rather than the same data used during tuning.
Put the risk beside the result
A performance improvement means more when the failure rate is shown right alongside it.
Make every claim testable
We want to be able to say exactly what would prove a result wrong.
Ask harder questions.
We are building models for checking evidence, modeling cells, screening sources, and learning over time. If you are working on similar problems, we would like to hear from you.