The ultimate guide to RL environments: building and scaling them in the LLM era
📝
281
Building and scaling RL environments for LLM training
Dr. Animesh Sinha Roy PRO
animesh1986
AI & ML interests
Senior Post Doctoral Research Fellow @ Raman Research Institute, Bengaluru ·
🔬 AI & ML Interests : AI agent developer · LLM agents with verified tools (LangGraph, smolagents) · quantum information physicist
Recent Activity
repliedto their post 1 day ago
Update to my Quantum Information Agent 🔬: it’s now an MCP server.
In my last post I introduced the agent: you ask in plain English, and tested physics tools do every calculation. https://huggingface.co/posts/animesh1986/853410084524560
Now the same 32 tools can be called directly from Claude, Cursor or your own agent: polarization-entangled photon states, local and global noise channels, Lindblad evolution, wave plates, maximum-likelihood tomography with error bars, certification of entanglement, steering and CHSH violation, and plots that come back as images.
1. No language model runs on the server, so every number comes from tested NumPy/SciPy code (now 678 tests and 30 analytic checks).
2. Each client gets its own session of stored states.
3. The Chat and Workbench tabs work as before.
Try it in Claude: add the MCP address below as a custom connector and ask “Make a Werner state with p = 0.8 and certify its entanglement”. Claude creates the state, runs the certification on the Space and explains the result. The picture shows that answer.
🔌 MCP address: https://animesh1986-quantum-information-agent.hf.space/gradio_api/mcp/
🔗 Demo: animesh1986/quantum-information-agent (setup for Claude and Cursor is under “Use these tools from your own agent”)
💻 Code and tests: https://github.com/animeshQKD/quantum-information-agent updated a Space 1 day ago
animesh1986/quantum-information-agent repliedto their post 3 days ago
Update to my Quantum Information Agent 🔬: it’s now an MCP server.
In my last post I introduced the agent: you ask in plain English, and tested physics tools do every calculation. https://huggingface.co/posts/animesh1986/853410084524560
Now the same 32 tools can be called directly from Claude, Cursor or your own agent: polarization-entangled photon states, local and global noise channels, Lindblad evolution, wave plates, maximum-likelihood tomography with error bars, certification of entanglement, steering and CHSH violation, and plots that come back as images.
1. No language model runs on the server, so every number comes from tested NumPy/SciPy code (now 678 tests and 30 analytic checks).
2. Each client gets its own session of stored states.
3. The Chat and Workbench tabs work as before.
Try it in Claude: add the MCP address below as a custom connector and ask “Make a Werner state with p = 0.8 and certify its entanglement”. Claude creates the state, runs the certification on the Space and explains the result. The picture shows that answer.
🔌 MCP address: https://animesh1986-quantum-information-agent.hf.space/gradio_api/mcp/
🔗 Demo: animesh1986/quantum-information-agent (setup for Claude and Cursor is under “Use these tools from your own agent”)
💻 Code and tests: https://github.com/animeshQKD/quantum-information-agentOrganizations
None yet