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
- When Assurance Work Displaces the Task examines overdefensiveness and overengineering in 35 episodes from 23 Codex conversations. The case study distinguishes useful assurance from added work that displaces the result. Case-study PDF · Case-study DOI
- Agent → Obsidian is a general method for turning AI agent work into linked Obsidian notes, with complete records, daily syntheses, and reusable methods. Includes a Codex example.
Papers and technical reports
- When Assurance Work Displaces the Task (2026, preprint). A retrospective, AI-assisted case study of overdefensiveness and overengineering, including local corrections, changing authorization, and necessary safeguards. Assurance-work PDF · Assurance-work DOI
- 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.
-
Agent → Obsidian: a general workflow for archiving AI agent conversations and tool traces, writing daily syntheses, and maintaining linked project and method notes in an existing Obsidian vault. Uses each agent's available records and the vault's existing sync setup; includes a Codex example with a dependency-free collector.
-
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