Independent AI-safety research.

I build instruments for measuring model instability and experimental architectures for making internal mechanisms easier to inspect.

Profile

Luis Jaime Ledesma Pérez

Independent researcher running TwoQuarks out of Guadalajara — no lab, no cluster. The research question is whether measurable behavioral or internal changes can precede visible model failure. I study that question through black-box probes (PfV / Molecule, from API outputs) and white-box measurements (CRAD, from the residual stream), while Dopamine is the architecture branch, where representation, integration, memory, and transmission are implemented as separate mechanisms. Public preprints, an open prototype, a Python package on PyPI, an MCP tool, and a live playground make the work inspectable.

Development stack

What I build with

  • Python · PyTorch · NumPy
  • REST APIs · MCP servers · provider adapters
  • RAG systems · tool-calling agents · multi-agent orchestration
  • Azure AI Foundry · Copilot Studio
  • HuggingFace · Netlify · PyPI packaging
Open to

What I can do on your team

  • AI-safety research engineering roles
  • LLM evaluation and red-team-adjacent work
  • RAG and agentic system audits
  • TwoQuarks control layer on existing pipelines — no weight modification, no retraining
  • Research collaborations, fellowships, applied safety teams
Contact

Let's talk