← Projects Knowledge Scale

CellScout

AI Research Assistant for the AIVC Literature

Transforms academic papers into a queryable knowledge graph — ask natural-language questions, get cross-paper synthesis powered by a multi-agent RAG system.

Core Capabilities
Natural Language Query Cross-paper Synthesis Knowledge Graph Method Lineage Tracking
Architecture
Orchestrator Agent Ingestion Agent Analysis Agent Code Agent

The Problem

The AIVC field is growing faster than any individual researcher can track. Knowledge is fragmented across hundreds of papers spanning scRNA-seq biology, graph neural networks, variational inference, foundation models, and wet lab validation. There is no tool that lets you ask "which methods address batch effects in single-cell data, and how have they evolved?" and get a grounded, cross-paper answer.

The Architecture

CellScout is built on an Orchestrator + Worker multi-agent pattern. The Orchestrator decomposes research queries into subtasks and routes them to specialized agents. Unlike flat RAG systems, CellScout reasons over a knowledge graph — capturing method lineage, dataset usage, and open problems as typed edges between nodes.

Ingestion Agent

Reads papers and extracts structured knowledge — methods, datasets, core claims, code links — into typed graph nodes.

Analysis Agent

Traverses the knowledge graph to answer cross-paper questions. Finds method evolution chains, dataset overlaps, and research gaps.

Code Agent

Retrieves reference implementations from the knowledge base and adapts them to your pipeline on request.

Orchestrator

Decomposes queries, schedules agents, synthesizes results into a coherent research-grade response.

Knowledge Graph Schema

The graph captures biological knowledge structure — not just document chunks. Typed edges enable reasoning that flat vector search cannot do.

Node Types
paper method dataset concept open_problem
Edge Types
INTRODUCES IMPROVES_ON USES RELATED_TO HAS_PROBLEM

Why Not Just RAG?

Standard RAG retrieves document chunks by vector similarity — it cannot answer "how did method X evolve into method Y" or "which papers share dataset Z and what did they find differently?" The knowledge graph enables graph traversal queries that surface relationships invisible to embedding search. CellScout combines both: vector search for relevance, graph traversal for relational reasoning.

Stack

Agents
Anthropic Claude API Custom Orchestration LlamaIndex
Knowledge Graph
Neo4j NetworkX Vector Embeddings
Deployment
FastAPI HuggingFace Spaces Docker