I build AI agent systems, reliable backend infrastructure, and local-first software products.

My projects follow one recurring rule: a feature is not “done” just because the demo looks right. I prefer executable verification—real E2E flows, fault injection, packaged-app smoke tests, provider-exact usage when available, and explicit boundaries around what the evidence does and does not prove.

AI / Agent 工程 · 全栈基础设施 · 本地优先产品
我更关注“东西真的跑起来以后是否成立”，而不是只展示功能截图。

Local-first, evidence-first long-term project memory for coding agents.

• Current Truth, scoped retrieval, conflict / freshness handling, and cross-session recall.
• Minimum Sufficient Context (MSC): return the smallest useful memory set instead of replaying everything.
• Controlled evaluations cover hard restart, wrong-scope isolation, memory updates, context eviction, and no-memory baselines.
• Current public evidence includes 396 Python acceptance tests, 14/14 plugin hidden validations, and 13/14 strict mode protocols; failed and inconclusive runs remain documented instead of being relabeled as wins.

DeepSeek Harness evidence · Executable benchmark artifacts · Architecture

A full-stack sandbox payment system built around financial correctness and failure recovery.

• Next.js + NestJS + PostgreSQL + Redis / BullMQ.
• Stripe and PayPal provider adapters behind a framework-neutral payment boundary.
• Signed webhooks, idempotency, refunds, outbox, ledger, reconciliation, queue workers, OpenTelemetry, Prometheus, and JSON logging.
• Real database / Redis E2E tests and a ten-scenario failure lab cover timeout, replay, restart, concurrency, and transaction rollback.

Architecture and quick start

A local-first, evidence-constrained AgentOps / AIOps control plane and sealed fault lab.

• FastAPI + React + PostgreSQL + durable workers + local Qwen / OpenVINO paths.
• Explicit state machine, RAG, Plan IR, approval gates, tenant-bound short-lived capabilities, idempotent execution, compensation, and post-action verification.
• Docker validation covers PostgreSQL, multiple competing workers, Nginx, Prometheus, fault lab, and real HTTP acceptance.
• Hidden-oracle evaluation keeps expected answers outside the runtime input.

Project guide

The standalone DeepSeek Harness bundle that connects DSH to MemoryOS.

• Scoped project-memory tools and provider-exact usage collection.
• Read-only by default; optional explicit cross-session write profile.
• 23/23 packaged contract plus real Loader/HMR tests passed against the locked DSH target.
• Evaluation emphasis: distinguish “memory works” from “agent success improved”.

Tests, failures, fixes, and limits · 中文说明

A Windows local-first e-commerce image production workbench.

• Tauri 2 + Rust + React / TypeScript + SQLite.
• Product-document ingestion, evidence confirmation, image planning, local Qwen3-VL assistance, and export workflows.
• Rust / TypeScript / E2E verification and internal-preview packaging; cloud providers are disabled by default.

Project status

• CodeDrake — Windows desktop pet that grows with project code size; PyInstaller EXE build, 34 regression tests, packaged GUI smoke, and GitHub Actions artifact.
• Project Bootstrap — lightweight planning / governance Skill that turns one rough idea into three formal AI-development files with bounded scope and stress-tested rules.
• Product Evidence Guard — local product-fact extraction and conflict detection with source provenance, unit normalization, and human-confirmation-only outputs.
• SafeReclaim — auditable Windows disk-cleanup Skill with read-only discovery, tamper-evident plans, and explicit approval before deletion.
• 出门别忘 — mobile-first PWA / Capacitor checklist app with Android and iOS projects.

Across projects, I repeatedly work on:

• Agent systems: state, memory, retrieval, tool governance, human approval, evaluation.
• Backend reliability: queues, idempotency, transactions, outbox, reconciliation, observability, fault injection.
• Local-first AI: desktop runtimes, local models, privacy boundaries, deterministic fallbacks.
• Evidence-first delivery: source provenance, executable acceptance, failure preservation, honest claim boundaries.

• No generalized success-rate claim from one solved task.
• Fixture / mock results are labeled separately from real-model or real-runtime evidence.
• Failed, quarantined, and inconclusive runs stay visible when they are informative.
• “Tool returned success” is not treated as proof that the user-level task succeeded.
• AI-generated facts remain candidates until the relevant deterministic or human gate accepts them.

我主要做三类项目：AI / Agent 系统、可靠全栈基础设施、本地优先软件产品。
旗舰项目是 MemoryOS、PayFlow 和 Harbor AgentOps；同时也持续做能真正打包、运行和验收的桌面与小型产品。