Building production-grade AI agents that connect LLMs with modern software architecture

A local-first desktop agent workbench. It builds a robust, recoverable agent harness centered around visual chat, workspace tools, approvals, and observable run records. In the process, a standardized, reusable desktop application architecture was distilled.

• Designed and implemented the core agent scheduling environment (Harness layer), decoupling model calls, context assembly, tool execution, permissions, and event logs into highly cohesive, observable, and recoverable modules.
• Built a secure workspace tool ecosystem (covering file operations, search, shell execution, LSP checks, etc.), implementing permission interception strategies based on sandbox boundaries and secret path protection to mitigate LLM execution risks.
• Developed an intuitive visual interface for the workbench, integrating run graphs, timelines, tool call trees, diff previews, and an approval inbox. Supported retry/skip actions for failed calls and drill-down analysis of sub-agent traces.

A personal knowledge management and spaced repetition system featuring agent-assisted chat, an advanced retrieval backbone, graph-based context, and cross-platform rich-text editing.

• Designed an agent-driven architecture that safely executes business data mutations through sandboxed tool calls and approval workflows.
• Architected a multi-path retrieval engine combining LLM query rewriting, full-text search, and tag inverted indexes, enhanced by a precise LLM reranking strategy.
• Developed a knowledge graph engine using xyflow, converting references and tags into high-density context to improve multi-hop reasoning across extensive notes.

A Claude Code-style AI coding assistant incorporating session management, diverse tools, plan-and-execute workflows, and context compaction.

• Led the architectural design, coding standards, and initial plugin selection for the AI coding harness.
• Engineered the core session module to robustly manage context, state transitions, lifecycles, and data flow.
• Built an interactive TUI/CLI layer supporting streaming text output, syntax highlighting, and keyboard-driven commands.

A secure, cross-platform instant messaging client leveraging the Matrix protocol, Android Compose UI, Electron, and a robust Rust SDK.

• Deeply customized the Matrix rich-text parser to support invite-link resolution, text highlighting, and seamless cross-platform navigation.
• Rebuilt the application-level deep link dispatcher and integrated QR generation and scanning, streamlining friend addition and room joining flows.
• Designed a pluggable navigation architecture using Appyx, cleanly decoupling room lists, settings, and room creation modules.