Independent developer. I build memory and self-evolution infrastructure for AI coding agents.
Most agent tooling optimizes the model call. I care about what happens between calls — what the agent remembers, what it forgets, and whether it gets better at your codebase or just faster at guessing.
dsh-expert-panel — A DeepSeek Harness plugin that turns a stuck main model into a cross-model adversarial review. Parallel read-only panelists → cross-examination → final arbitration, with a hard output contract: you get an assignable work order, not a summary.
skill-evolution — Closed-loop skill self-evolution for coding agents. Post-turn review, directive fusion, four-gate safety validation, SHA-256 ledger rollback. Skills that change themselves need an undo button.
awesome-dsh-plugin — The curated plugin index for DeepSeek Harness. If a plugin exists and works, it's here.
Local-first, self-hosted, no vendor holds my data. Multi-agent memory runs on my own hardware over encrypted tunnels — not because it's fashionable, but because memory is the one asset worth owning outright.
I run several coding agents in parallel and compare their outputs rather than trusting one. Cheap models for routine work, expensive ones only when the decision actually matters.
I write about agent memory architecture, self-evolving tooling, and the unglamorous plumbing that makes agents reliable. Mostly in Chinese, sometimes in English.
X: @vesisjdodos
Open to conversations about memory architecture and agent tooling. Issues and PRs welcome on any repo above.