Medical AI that grades its own evidence.

Every answer carries its grade, links the studies it rests on and refuses when the evidence is thin.

Start freeSee it in use

Chat app

Consumer app, iOS

Ask it the way you would ask a colleague

  • Every answer carries its grade on the GRADE scale.
  • The study is attached and the citation opens the PubMed record.
  • It refuses when the evidence is thin and says why.
See it in use

Every answer is read from the primary literature

  • PubMed
  • PubMed Central
  • ClinicalTrials.gov
  • DailyMed
  • openFDA

Supported by

Potsdam Transfer Startup Service, University of Potsdam

See it in use.

The grade, the citations and the refusal travel with the answer into whatever interface it lands in. The chat app is ours; the protocol builder and the teaching tool show what can be built on the API, and the browser panel is a concept that doesn’t exist yet.

Ask it the way you’d ask a colleague

A chat interface your patients or clinicians already know. What arrives is not a confident paragraph: it is a graded answer with the study attached, and a refusal when there is nothing to stand on.

9:41

Ask

Can a patient on apixaban take St. John’s Wort for low mood?

Reading the label and three records

What dose of BPC-157 for a tendon then?

Searching for human dosing data

One small study and twenty good trials produce the same confident paragraph. So every answer here carries its grade, on the GRADE scale you already use.

This is what one answer looks like.

Not a summary with a link at the bottom. The grade sits on the answer, the studies are named, and every citation is one click from its record.

Question

Does berberine lower fasting glucose in type 2 diabetes?

Grade B: Moderate certainty

Randomised trials and meta-analyses report meaningful reductions in fasting glucose with berberine in type 2 diabetes. Results vary between studies and most follow-up is short.

An example answer in the shape the engine returns it. Every citation opens the PubMed record it names.

What a grade means

A  High

Well-run trials that agree with each other. You can lean on it.

B  Moderate

Good studies with limits in size or design. Read them before you act.

C  Low

Small, early or mixed studies. A reason to look closer, not a plan.

D  Very low

Thin or conflicting. Treat it as a hint and nothing more.

Evidence too thin

The literature can’t support an answer, so the engine refuses and gives the reason.

Refused

A safety rule fired. A personal dose for a compound with no label dosing is never given, whatever the studies say.

Four steps. The third one is ours.

Retrieving papers and attaching citations are solved engineering problems. Grading how much weight the evidence can carry is done by expert committees, over months. We automate that step.

Does creatine supplementation harm kidney function in healthy adults?

  • PubMed: 3 records
  • ClinicalTrials.gov: 0 registered trials matched
  • DailyMed: not a labelled drug

  1. Retrieve

    Your question goes to the primary literature: PubMed, PubMed Central, ClinicalTrials.gov, DailyMed and openFDA. No summaries of summaries.

  2. Ground

    Each sentence in the answer is tied to the study or label span it came from. A claim with no source doesn’t ship.

  3. Grade

    The engine weighs the design of each study, how many people were studied and whether the results agree, then grades the certainty from high to very low. The grade sits next to the answer, not in a footnote.

    Starts atHigh
    Imprecisionone step down

    The controlled trials are small, so together they enrolled too few people to settle the question.

    Lands atModerate

    When a grade drops a level, the reason is stated.

  4. Refuse

    When the sources can’t support a safe answer, you get a refusal and the reason. A gap in the data stays a gap.

Built for the rooms where the answer matters.

Clinics

A patient asks about a supplement mid-consult. Get a graded, cited answer before the visit ends and show them the study.

For clinics

Pharmacies and integrative practices

Run everything a patient takes through pairwise checks. Interactions come with their label text. Unknowns get flagged, not passed.

See the product

Research and medical affairs

See what the evidence says before you commit weeks to reading it. Every claim points back to its record.

Read the method

Developers

Put graded, cited answers behind your own product over one API, with refusal semantics your compliance team can reason about.

For developers

Sometimes the most useful answer is no.

A tool that always answers is a tool you can’t trust. When the evidence comes only from animals, ours answers with a clear note that says so, and it holds back a dose to take for any compound without label dosing.

Question

Optimal peptide dose for tendon repair in humans?

Evidence too thin

Human dosing evidence is insufficient. The available studies are animal models and case reports, which can’t support a dose recommendation.

For developers

The same engine, over plain HTTP.

One authenticated call returns the answer, the claims, the citation behind each claim and the grade. A refusal comes back as a normal reply, because it is one. Twenty-three public endpoints, documented on one page.

Read the API referenceWhat you can build
curl "https://novolab.io/v1/ask" \
  -H "X-API-Key: $NOVOLAB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "question": "Does creatine harm the kidneys?" }'

Ask it something you already know the answer to.

The fastest way to trust a grading engine is to watch it grade a question you can check yourself.

Start freeTalk to us