Useful AI needs more than a strong model response. It needs continuity, evidence, recovery, safe boundaries, and a product surface people can trust.

My projects explore that full stack:

• Agent reliability and interoperability — session handoff, task recovery, context portability, idempotent continuation
• Evidence-grounded AI — local analytics, traceable SQL, answer validation, citations, persistent task queues
• Native AI products — privacy-aware macOS experiences built with Swift, SwiftUI, and AppKit
• Developer tools — Python CLIs, skills, adapters, automation, tests, and multi-version CI

Cross-agent task continuity for Claude Code and Codex. When one coding agent hits a usage limit, the CLI extracts portable local session context, reconciles Git state, redacts common secrets, and launches the other agent with a safe takeover contract—without requiring the limited model to respond.

Python · Agent interoperability · Session recovery · Developer tools

A local-first AI analytics workbench for product, operations, and business teams. It combines spreadsheet ingestion, DuckDB analysis, executable SQL evidence, answer validation, user-correction memory, persistent AI task queues, and report generation.

Swift · DuckDB · Evidence-grounded AI · Analytics agents

A native macOS work-memory system that turns window, web, typing, OCR, and document context into searchable sessions, grounded answers, summaries, projects, and action items—while keeping collection and storage local-first.

SwiftUI · Local retrieval · Knowledge systems · Privacy

A competitive-intelligence monitor that brings together RSS, web pages, Tavily search, and Chrome session sources, then applies AI-assisted deduplication, synthesis, trend analysis, and report generation.

Swift · Information pipelines · AI synthesis · Product intelligence

A native, privacy-first macOS menu-bar utility for safe daily cleanup and disk-health visibility. It reflects the other half of my work: turning systems concerns into focused products that people can understand and control.

SwiftUI · AppKit · macOS · Product engineering

Continuity over fragile sessions.     让任务跨越会话和工具继续运行。
Evidence over plausible answers.      让结论能够追溯、验证和复现。
Local-first when privacy matters.     敏感数据优先在本地处理。
Safe recovery over blind retries.     恢复之前先核实，避免重复副作用。
Working products over AI demos.       把模型能力做成真正可用的产品。

I am currently focused on:

• cross-agent continuity, memory, and task recovery;
• reliable tool-using agents with observable and testable behavior;
• evidence-grounded analytics and decision-support systems;
• native interfaces for privacy-sensitive AI workflows.

Python · Swift · SwiftUI · AppKit · DuckDB · SQLite · macOS · Agent tooling · Local-first AI · Automation · CI