BioTeam-AI: Personal AI Science Team for Biology Research

JangKeun Kim  |  Weill Cornell Medicine
GitHub | Live Demo
Problem

Biology researchers spend 8–10 hours per week on manual literature review across fragmented sources. No unified tool exists to automatically collect, deduplicate, score, and summarize papers from PubMed, bioRxiv, arXiv, GitHub, HuggingFace, and Semantic Scholar.

Solution

BioTeam-AI is a multi-agent system powered by the Claude API that automates the full research intelligence pipeline. 18 specialized LLM agents handle literature monitoring, hypothesis generation, experiment design, and knowledge management through a unified dashboard.

Architecture
Tier 1
Directors
Research Director Knowledge Manager Project Manager
▼   orchestrate   ▼
Tier 2
Teams
Literature Scout Data Analyst Methodology Advisor Hypothesis Generator Experiment Designer Peer Reviewer Compliance Officer Writing Coach Collaboration Facilitator Tool Specialist
▼   delegate   ▼
Tier 3
Engines
Digest Engine Ambiguity Engine RCMXT Scoring Citation Validator Session Manifest
▼   powered by   ▼
Infra
FastAPI SQLite + SQLModel ChromaDB Next.js 16 SSE Real-Time Docker Compose
Key Metrics
18
LLM Agents
across 3 Tiers
6
Data Sources
Integrated
725+
Automated
Tests
$0.005
Per Digest Run
(vs $50 manual)
962x
ROI on
LLM Cost
4.2s
Avg Pipeline
Execution
69
Source Files
100% Type-Safe
100%
Standalone
No Vendor Lock
Tech Stack
Python 3.12 FastAPI Anthropic Claude API Haiku / Sonnet Instructor SQLModel ChromaDB Next.js 16 Tailwind CSS shadcn/ui SSE Docker Compose
Why This Matters for AI in Science

This project is a practical demonstration that Claude can serve as the backbone of a full scientific research workflow — not as a chatbot, but as a coordinated team of specialized agents that drastically reduce the cost and time of literature intelligence. By achieving a 962x ROI at $0.005 per digest, it proves that AI-driven research automation is economically viable for individual labs. BioTeam-AI bridges the gap between frontier AI capabilities and everyday scientific practice, directly aligned with Anthropic's mission to make AI genuinely useful for advancing science.

Key Features
  • Multi-source literature digest with word-boundary relevance scoring across PubMed, bioRxiv, arXiv, GitHub, HuggingFace, and Semantic Scholar
  • RCMXT evidence scoring: 5-dimension assessment (Reproducibility, Condition Specificity, Methodological Robustness, Cross-Omics Consistency, Temporal Stability)
  • Ambiguity detection engine with automated resolution workflows for contradictory findings
  • Real-time SSE activity feed with live agent status monitoring and streaming responses
  • Negative results knowledge base with full CRUD operations and semantic search
  • Direct query streaming interface with persistent conversation history and context retention