I build evidence-driven infrastructure for AI agents: memory that survives lossy context, completion claims bound to current state, and reusable skills that can be tested and operated.
My work sits between agent research, infrastructure engineering, and hands-on product building. I prefer public artifacts, repeatable checks, bounded experiments, and honest negative results over convincing demos.
Technical Staff of Myself.
Email · X / Twitter · Hugging Face
- BeforeDone binds coding-agent completion claims to fresh verification evidence from the repository state being described. BeforeDone technical report · BeforeDone DOI
- Agent Memory System is a local-first, cross-agent memory layer that keeps raw evidence local and promotes only reviewed, redacted, portable memory. Agent Memory live site · Agent Memory on Hugging Face
- Governed Skill Evolution compares no Wiki, flat history, and persistent Wiki in a prospective ablation and cross-model transfer study. Persistence reduced input tokens and final-rule length in this setting, but did not improve mean task quality. Governed Skill Evolution PDF · Governed Skill Evolution DOI
Papers and technical reports
- Governed Skill Evolution from Persistent Agent Experience (2026, preprint). A prospective ablation and cross-model transfer study of no Wiki, flat history, and persistent Wiki. GSE PDF · GSE DOI
- BeforeDone: Fresh, State-Bound Evidence for Coding-Agent Completion (2026, technical report). A frozen, reproducible study of whether completion claims carry passing verifier evidence bound to the repository state they describe. BeforeDone DOI
- Thin Harness, Strong Contracts (2026, preprint). A production-oriented study of agent harnesses as runtime, permission, state, replay, memory, evaluation, audit, and human-review infrastructure. Thin Harness DOI
- SkillOps: A Practical Framework for Designing, Testing, and Operating Modular Skills in Personal AI Agents (2026, paper and software artifact). A lifecycle framework for Skill contracts, tests, evaluation, versioning, safety, and operations. SkillOps PDF · SkillOps concept DOI
Research and applied systems
-
Intent Formation: an experimental Codex plugin for following changing goals through small samples and feedback. Includes opt-in, sourced state and recovery, a DeepSeek Harness adapter, and Windows/Linux/macOS CI. The final synthetic confirmation was incomplete and missed its match-improvement gate; v0.3 is a public research snapshot, not a released product. Current goals and status · Evidence and release decision
-
Agent Run Evidence Toolkit: independent repositories for retaining observable runs, evolving instructions, creating resettable scored workflow Cases, and exchanging portable evidence. Runtime Evolution Workbench · Workflow Environment Factory · RunCase Interchange · DeepSeek RunCase adapter
-
Agent Memory System: local-first, cross-agent memory with reviewed, redacted, portable outputs. Agent Memory live site · Agent Memory on Hugging Face
-
Context Continuity: preserves task invariants across compaction, resume, and handoff without treating reconstructed context as user authority.
-
Completion and execution reliability: fresh, file-bound completion evidence plus contract checks before high-impact actions. BeforeDone · Execution Fidelity Guard · DeepSeek execution adapter
-
Skill lifecycle and safety: an evidence-gated path from selected runs to one reviewable Skill change, paired with static scanning for risky Skill content and archives. Experience to Skill · Skill Security Guard
-
Research Toolkit / 研究工具箱: methods, workflows, and checks for AI-assisted research reports, from questions and source analysis to section-by-section writing, review, and revision. Project guide
-
Editorial Guard: Chinese style editing and delivery-language validation with shared fidelity checks, two project Skills, and two model-annotated benchmarks. Disagreements and negative results are retained; the two-family core comparison does not establish general editing benefits. Project site · Hugging Face
Working approach: Bind important claims to observable evidence. Use deterministic checks for deterministic behavior and model evaluations where behavior is stochastic. Keep sensitive raw evidence local; publish reviewed, redacted, reproducible artifacts, including real failures and user corrections.
Stack: Go · Python · TypeScript · JavaScript · Shell · HTML/CSS