An Agent Engineer with a strong foundation in software architecture. I specialize in integrating Large Language Models (LLMs) into real-world workflows to build efficient, stable AI agent products and cross-platform clients. With deep expertise in Agent Skills, Memory, and RAG, combined with solid engineering experience in desktop and frontend architecture (Vite+, Electron, Rust), I focus on building observable, recoverable systems while maintaining a high delivery rhythm.
I
AI Engineering
A complete engineering loop across Skills, Harness, Memory, and RAG
Encapsulating model capabilities into Skills, connecting external tools via MCP, and managing session memory in an observable harness — the foundation for shipping reliable, production-grade agents.
I
Agent Skills & Tools
Encapsulate domain actions into reusable agent skills and connect external capabilities via the Model Context Protocol (MCP). Design tool-calling layers with approval workflows, structured output, and streaming interactions, leveraging sub-agents to break down complex tasks.
II
Agent Harness & Architecture
Treat the harness as the product layer that decides how much of a model's capability actually ships — the same engineering path Anthropic (Claude Code) and Cursor bet on. Build production-ready agent runtimes covering session management, state-machine scheduling, lifecycle hooks, approval-gated tool execution with per-profile allowlists, event sourcing, and automatic recovery of interrupted runs.
III
Memory & Context Engineering
Manage LLM contexts efficiently. Utilize short-term memory, graph-based context, prompt caching, and token budget control to balance multi-hop reasoning quality with computational costs.
IV
RAG & Retrieval
Construct multi-dimensional knowledge retrieval pipelines. Combine text chunking, vector embeddings, full-text search, and inverted indexes, optimized with LLM reranking strategies to ensure the accuracy and traceability of generated content.
II
Selected Work
Selected Work & Engineering Delivery
A curated selection of production-proven systems, spanning agentic knowledge bases, AI coding tools, encrypted communication clients, and desktop trading systems.
A multi-agent decision workflow for hotel procurement, covering the full loop of bid collection, review scoring, contract review and archival audit. Owned the full stack — agent runtime, backend services and the React workbench — turning expert-grade review methodology into an auditable, replayable automated decision process.
Designed a collaboration architecture of a supervisor agent plus multiple specialist review sub-agents, each owning one of the technical, commercial and contract-compliance review chains.
Worked with domain professors to encode supply-chain, negotiation and legal-compliance review methodology into each sub-agent's scoring rubric and expert-role persona, so every sub-agent reviews from an explainable, discipline-specific perspective.
Built a Slock-style 'group-chat' collaboration layer for the sub-agents: each is a persistent member with its own identity and memory that joins a shared channel, exchanges structured messages over a shared context, and is coordinated by the supervisor through turn-taking and @-mention handoffs — so the specialist agents deliberate over a single bid in one room and converge on a review verdict.
A local-first desktop agent workbench built around a robust, recoverable agent harness — visual chat, workspace tools, approvals, observable run records — from which I distilled a standardized, reusable desktop application architecture.
Designed and implemented the core agent runtime (harness layer), separating model calls, context assembly, tool execution, approvals, sub-agent delegation, and event logging into observable, recoverable modules, with multi-provider support (Anthropic, OpenAI, AI Gateway — including an API-mode toggle and a proxy-aware network layer).
Deliberately narrowed the workspace tool surface to a minimal, auditable set (read / ls / grep / edit / write, with writes gated by approval), and paired it with secret-path filtering and secret redaction before any tool data is persisted (API keys and tokens auto-scrubbed), keeping execution risk contained and auditable.
Built the Run Inspector and a durable pending-approvals list: each assistant message opens that run's tool calls and event timeline, and pending approvals resume across app restarts (auto-expiring after 7 days).
Vite+ · Electron · React · Hono · AI SDK · SQLite · Rust CLI
III
Internal knowledge agentic system
Agent Engineer · 2025.01 - present
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 a Mastra agent-driven architecture that safely executes business-data mutations through sandboxed tool calls with approval gates (data-mutating tools require approval; read-only tools stay ungated).
Architected a multi-path retrieval engine combining LLM query rewriting, full-text search, title match, tag inverted indexes and pgvector embeddings, 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.