LabMind
Live Agent: https://labmind-840267766274.us-central1.run.app/
Demo Video: https://youtu.be/_BhIjSgo9Jo
An AI agent dashboard for autonomous scientific experiment optimization. LabMind demonstrates a human-in-the-loop workflow where an AI proposes the next batch of laboratory experiments, researchers review and optionally adjust the proposal, and then approve it — creating a continuous optimization loop.
Research domain: mRNA-LNP (lipid nanoparticle) delivery optimization — finding the optimal combination of pH, temperature, mRNA concentration, lipid ratio, and incubation time to maximize transfection efficiency.
What This Agent Does
LabMind runs a LangChain ReAct agent powered by Claude that closes the loop between experimental results and next-step decisions.
-
Process experimental data
The system starts by processing fluorescence microscopy TIF images from a batch of experiments. Using computer vision methods(Otsu thresholding and watershed segmentation), it measures transfection efficiency for each condition.
-
Identify the best-performing result
Once all experiments in the batch are analyzed, the agent compares outcomes and identifies the top-performing parameter combination.
-
Generate the next experiment proposal
The agent uses the result as the center point for Bayesian optimization, generating a new set of parameters for the next batch.
-
Explain the outcome
Alongside the proposal, the agent produces a structured scientific analysis that explains what drove the result and why the next range is worth exploring.
-
Refine with human input
Before moving forward, the researcher can set constraints in the chat interface. The agent incorporates these constraints, regenerates the parameter space, and updates both the proposal and the reasoning.
This system is designed for real laboratory settings where experiments run in parallel and each batch takes hours or days to complete. The agent handles the optimization loop, while the researcher remains in control of the final decision.
Where to use it: This serves as a starting point for lab workflows that require iterative optimization with human oversight, including drug formulation, cell culture condition screening, and other batch-based experimental systems.
Tech Stack
| Layer | Technology |
|---|
| Frontend | Next.js 15 + TypeScript + Tailwind CSS + Recharts |
| Backend | Python 3.11 + FastAPI + LangChain |
| LLM | Claude API (claude-sonnet-4-20250514) via Anthropic SDK |
| Agent framework | LangChain ReAct agent with AgentExecutor + ConversationBufferMemory |
| Optimization | Latin Hypercube Sampling (scipy.stats.qmc) |
| Image processing | tifffile, scikit-image (Otsu threshold, watershed), Pillow |
| Data store | JSON files (no database) |
| Deployment | One Google Cloud Run container (FastAPI serves the exported Next.js UI), deployed by GitHub Actions on every push to main |
| Image dataset | BBBC016 (Broad Institute) — GFP fluorescence TIF files |
How to Run
You need two terminals: one for the backend, one for the frontend.
Prerequisites
- Python 3.11
- Node.js 18+
- An Anthropic API key
Backend
cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Edit .env and set ANTHROPIC_API_KEY=sk-ant-...
uvicorn main:app --reload --port 8000
The backend starts at http://localhost:8000. On first startup it resets any in-flight state to IDLE and clears processed image PNGs (they are regenerated on each run).
Frontend
cd frontend
npm install
cp .env.local.example .env.local
# .env.local.example already points to localhost:8000 — no changes needed
npm run dev
The frontend starts at http://localhost:3000.
Verify
Open http://localhost:3000 to check the Welcome page.
Project Structure
LabMind/
├── backend/
│ ├── main.py # FastAPI app, all endpoints, CORS, background tasks
│ ├── agent.py # LangChain ReAct agent, tool loop, chat
│ ├── tools.py # 4 LangChain tools (load, rank, optimize, images)
│ ├── image_processing.py # TIF reading, Otsu threshold, transfection rate computation
│ ├── state_manager.py # state.json read/write with thread locking
│ ├── models.py # Pydantic request/response models
│ ├── requirements.txt
│ ├── data/
│ │ ├── state.json # global state machine state
│ │ ├── batches/ # batch_B1.json, batch_B2.json (pre-seeded), batch_B3+.json (generated)
│ │ └── proposals/ # pending.json (agent-generated, overwritten on regenerate)
│ └── static/images/ # BBBC016 TIF files + processed/ GFP PNGs (runtime-generated)
│
└── frontend/
└── src/
├── app/ # Next.js routes: /, /history, /experiments, /analysis
├── components/ # WelcomePage, RunningView, AgentAnalysis, Sidebar, charts, etc.
├── hooks/ # usePolling (4s), useChat
└── lib/ # api.ts (typed fetch client), types.ts
References
