01 — The problem
A research team works one question for weeks. It cannot build on an answer it cannot audit. Today it meets three problems.
02 — Who it is for
Researchers at universities and institutes, scientists at small and mid-size pharma and biotech companies, and the informatics teams who choose their tools.
Universities and institutes
Small and mid-size companies
The people who choose the tools
Partner, investor or future colleague? Read about Amordad
03 — Our approach
A controlled agent is an AI agent that follows a written plan, uses a short, fixed list of tools, and may cite only evidence the system collected.
1 · Question
You ask
A research question in plain English.
2 · Agent
It follows a plan
A written research plan, a fixed set of tools, and a budget for steps and time.
3 · Sources
It queries the data
A knowledge graph for relations, full-text papers for supporting passages.
4 · Evidence
The server numbers it
The server, not the AI, collects each result and gives it a number.
5 · Answer
You check it
Each claim carries a number that opens its source. Gaps are stated.
04 — Products
Each product applies controlled agents to one kind of research task. Cognigraph is live today.
Product 1 · Cognigraph · live
Cognigraph is an evidence-first research assistant for biomedicine. Ask about genes, diseases, drugs or tissues in plain English. Cognigraph searches a 137-million-edge knowledge graph and 3.6 million full-text papers, then answers with named identifiers and numbered citations you can open. If the evidence is not there, it tells you.
Links in the graph · as of October 2026
{{ schema.cur.count }}
{{ schema.cur.name }}
{{ schema.cur.text }}
Example question
{{ schema.cur.q }}
Graph reasoning in small, checkable steps
ground
A name becomes a standard identifier.
expand
A node opens to its neighbours.
connect
Find the links between two nodes, over several hops.
combine
Join sets, such as the genes two diseases share.
rank
Order the results, with the evidence for each.
Every graph step is a defined query with an exact result, not free AI text. Large lists pass by reference, not through the AI. The graph runs in a native parallel graph database for multi-hop queries.
Knowledge graph · relations
137.5M
edges · 53 sources · about 10 million nodes
The ROBOKOP knowledge graph 1.0.2 from RENCI, on the Biolink Model 4.4.2. Some of its sources need permission for commercial use.
Full-text papers · supporting evidence
3.6M
papers · about 134 million passages
The PubMed Central open-access subset from the U.S. National Library of Medicine. Each article keeps its own licence.
All counts as of October 2026.
Cognigraph · how an answer is built
Cognigraph follows a written plan. The server collects the evidence. The AI writes the answer, but it may only cite what the server collected.
Is TCF7L2 linked to type 2 diabetes, and where is it expressed?
Illustration · examplePaper passage · full text
Paper passage · full text
In short
The graph links TCF7L2 to type 2 diabetes , and papers discuss the same link . The graph links TCF7L2 to the pancreas .
Nothing found. The graph returned nothing for TCF7L2 in kidney.
Point at a number to see its evidence in the drawing.
Product 2 · Target Investigator · coming soon
Target Investigator is a guided workflow inside Cognigraph for drug-target questions. It is in testing now.
01
Define diseases
Up to four.
02
Review candidates
Candidate genes, ranked.
03
Open a target
Disease links, biology, tissues and drugs.
04
Evidence map
Conflicts and gaps in view.
05
Copy the report
Ready for your notes.
A side drawer opens the source of any citation. We will give a release date when it is ready.
Product 3 · Chemical synthesis · [STATUS]
A platform for chemical synthesis research. It applies the same controlled-agent method to finding and comparing ways to make a compound.
Work back from a target compound to possible starting materials.
Explore the steps of possible synthesis pathways.
Evaluate and compare potential routes.
[DETAILS AND DATA SOURCES TO CONFIRM BEFORE LAUNCH]
05 — Technology
Rules in code, not in prompts. The products do not trust the AI to behave. The system around it enforces the rules.
Written research plans and a short, fixed list of tools. Ten tested plans for common questions. No web browsing.
The server, not the AI, collects and numbers the evidence. The AI can only cite what the system retrieved.
BM25 and semantic search over about 134 million passages, merged by reciprocal rank fusion.
Graph links checked against the papers, with exact quotes confirmed in code.
Limits on steps, tool calls and time. At the limit, the answer is marked as partial, not hidden.
A live run trail, a debug view of each call, and export as a web page or a transcript.
06 — Questions and answers
Request access
That is the fastest way to judge a research tool. Ask for access and test Cognigraph on your own field.
What you get
We start with non-commercial research access. [ACCESS TERMS]
Prefer to talk first? Book a demo.
We will reply to your work email within [REPLY TIME]. While you wait, walk through how Cognigraph builds an answer.
01 — Mission
Our aim is simple. Every answer should be something a researcher can check: the source named, the evidence numbered, and the gaps stated.
02 — How we work
Every cited claim opens its source.
Nothing found is an answer, not a gap to fill.
The system enforces the rules. We do not ask the AI to behave.
03 — Progress
For partners, investors and future colleagues. Each product shows its current status.
An evidence-first research assistant for biomedical questions. It answers from a knowledge graph of 137.5 million curated edges and 3.6 million full-text papers, as of October 2026.
Target discovery: a guided workflow inside Cognigraph for drug-target questions, in testing now.
Retrosynthetic analysis, synthesis pathway exploration and the evaluation of possible routes.
04 — Team
[NAME]
[ROLE]
[NAME]
[ROLE]
[NAME]
[ROLE]
[NAME]
[ROLE]
05 — Company
06 — Data credits
Cognigraph builds on the ROBOKOP knowledge graph from RENCI and the PubMed Central open-access subset from the U.S. National Library of Medicine.
For research use, not clinical decisions.
Legal
Last updated [DATE] · {{ langNote }}
{{ l }}