multi-turn inc.
INDEX

index

AI Scientist

product

Preview — In Development

Most scientific knowledge is trapped in PDFs.

Think about what it takes to properly digest a single paper. Triaging the literature costs hours per paper, setting up a reproduction environment takes days, and tracing a citation network by hand costs hours again. Work like this is closer to structured information processing than to deep intellectual judgment. The moments that call for judgment sit elsewhere; the labor in front of them is exactly what an agent can handle well.

So it starts by compiling papers into a machine-readable structure. The goal is to cut context cost to less than half of the original while preserving source information without loss.

{
  "paper_id": "arxiv:2405.21060",
  "title": "Transformers are SSMs",
  
  "claims": [{
    "id": "C1",
    "statement": "Mamba-2 achieves 8× throughput",
    "evidence": ["E1", "E2"],
    "confidence": 0.95
  }],
  
  "evidence": [{
    "id": "E1",
    "type": "table",
    "location": "Table 3",
    "reproducible": true
  }],
  
  "methods": [{
    "id": "M1",
    "description": "Structured State Space Duality",
    "artifacts": ["A1"]
  }],
  
  "artifacts": [{
    "id": "A1",
    "type": "code",
    "url": "github.com/state-spaces/mamba",
    "verified": true
  }]
}

It's a graph where claims, evidence, methods, and artifacts reference one another. Not the paper's sentences but this structure becomes the raw material for every stage that follows.

The system isn't a single model. It composes several foundation models as sub-agents matched to the character of each task. The place that analyzes complex argument structure, the place that generates and runs code, and the place that processes a long paper in full are all different, and into each place goes the best frontier model of the moment. Models are replaceable parts; the structure is the machine.

On top of that structure, a Supervisor distributes work dynamically and synthesizes the results. Beneath it work an ingestion agent that parses PDFs and URLs, a citation agent that searches arXiv, Crossref, Semantic Scholar, and PubMed, a compiler agent that transforms papers into the schema above, a verification runner that executes code in Docker to compare metrics, and a critic agent that judges quality. The final approval is done by a human.

The execution environment is built on Docker, Jupyter, and pytest, and data is stored in a JSON Research Object DB and vector indices. Orchestration uses LangGraph with custom routing.

Right now each stage of the pipeline works independently. Paper structuring and citation graph construction have reached production level. Autonomous experiment reproduction and extension discovery are in development, and reproducing a paper without code from its method description alone is still at the research stage. Full autonomous orchestration is being stabilized.

AI Scientist | Multi-turn Inc.