Accountable AI infrastructure for human stakes work.
Clinical duty of care, translated into systems that can be inspected, corrected, constrained, and held accountable.
I build local first memory, agent safety tooling, career decision safety, and research workflows for situations where correctness, consent, and auditability matter.
Flagship Work
Four public projects define the current research software arc, accountable memory, career decision safety, social science workflows, and risk scaled agent execution.
mneme
Local first, audit safe memory for Claude Code and MCP clients. Markdown stays the source of truth.
Decision safetyVocationOS
Human supervised career decision safety with claim graphs, high stakes gates, and append only audit.
Research workflowsClaude Code for Social Scientists
Bilingual research guide for agentic coding without abandoning method, citation, or disclosure.
Agent executionmergen
Risk scaled execution harness for AI coding agents, workflow orchestration, and adversarial verification.
Operating Thesis
AI systems that touch human decisions need more than fluent output. They need evidence discipline, correction paths, consent boundaries, and audit trails.
Inspectable before intelligent
Memory and recommendations should be traceable before they become persuasive.
Automation with brakes
Drafting, deciding, and acting are different risk classes. Human authorization is a system boundary.
Evidence over performance theatre
Weak evidence should not become precise public claims just because the interface is confident.
Local first when stakes are human
Private reasoning surfaces need redaction, reversibility, and operator control by default.
Writing and Collaboration
Current writing focuses on accountable AI memory, career decision safety, agentic research workflows, and clinical duty of care as an engineering principle.