NovusGraph
Research literature is a graph. We treat it like one.
Papers were designed for print — one long linear document per idea, citations as numeric stubs, evidence buried in figures. NovusGraph pulls that structure inside out: every paper becomes a set of atomic claims, each backed by evidence atoms, each linked to the method families and implementation views it participates in. Then you query the graph.
The problem we care about
Reading a single paper thoroughly takes hours. Reading twenty in a subfield takes weeks. By the time you finish, the field has moved. And the parts you actually need — the specific claim, the specific ablation, the specific method comparison — are locked inside PDFs that don't talk to each other.
Traditional literature-review tools stop at retrieval: they hand you back a list of PDFs. NovusGraph goes further. We extract the structure inside the paper — claims, evidence, methods, families, implementations — and store it as a graph. Once the graph exists, the interesting questions become searches on the graph, not trawls through prose.
Our approach in three layers
Claims and evidence atoms
Every claim in a paper — "our method beats X on Y," "we hypothesize Z," "the loss converges after N steps" — gets extracted as a first-class node. Every claim is anchored to the specific evidence atoms that support it: a table row, a figure caption, a paragraph. When you ask "what does the graph say about method X on benchmark Y," you get back the actual atoms with their sources, not a summary that could be paraphrased away.
Method families and implementation views
Papers within a subfield share method families (attention variants, distillation strategies, VAE reformulations). NovusGraph identifies these families and links papers into them, so you can see the family history and reach the neighborhood of any technique without knowing its exact name.
Ask Graph, Paper Chat, Trace
Three surfaces on top of the graph: Ask Graph for cross-paper questions, Paper Chat for a single paper you're actively reading, and Trace for verifying a specific claim end-to-end down to the source. Everything is grounded — no answer without cited atoms.
Who we are
A small research team building the tools we wished we had while doing literature reviews. We work quietly, ship weekly, and read a lot of papers so you can read the ones that matter.
We are in early preview. If you want to reach us about a use case, a partnership, or a bug, the fastest path is through the workspace — the assistant will route the message.