Etcetera

Etcetera

AI / Agent Engineer

Summary

Agentic full-stack engineer with 4 years of shipping experience across Agent Skills, Harness, Memory, RAG and cross-platform clients. I focus on AI software systems that are observable, recoverable and continuously iterated, not one-off prompt demos.

Core Skills

Agent Skills/tools
Wrap domain actions with Skills / MCP / Tool Calling, then add approval gates, structured output, streaming feedback and failure recovery.
Harness/runtime
The model sets the ceiling; the harness decides how much of it ships — the Anthropic / Cursor thesis. Built that layer hands-on in Etyon and the AI coding agent: session lifecycle, plan / exec, compaction, approval gating, sub-agent delegation and event-sourced recovery.
Memory/context
Long-term memory without per-turn full retrieval — landed in Etyon as a resident project digest plus on-demand memory tools; paired with graph context, token budgets, prompt caching and wasTruncated markers to manage input quality and reasoning cost.
RAG/retrieval
Chunking, embeddings, FTS, title matching, tag indexes, hybrid recall, LLM reranking and citation traceability.
Product/engineering
TypeScript, React, Svelte, Electron, Tauri, Expo, Hono, oRPC, PostgreSQL, Rust and Docker across frontend, client and server surfaces.
Python/AI / tooling
Modern Python tooling with uv / Ruff; comfortable with FastAPI / Pydantic, hands-on RAG / embeddings via LangChain, and a Polars / pandas / NumPy data stack with Plotly reporting.

Experience2022.10 - present · agentic systems, cross-platform clients, engineering velocity

Beijing Shenzhou Yunfu Technology Agent Engineer
Hangzhou Juhe Liandong Technology Frontend / Rust Engineer
Suzhou Liangtang Information Technology Frontend / Blockchain Engineer
Shanghai Bingjiang Information Technology Frontend Engineer

Projects

DCS procurement AI workflow

· procurement AI workflow workflow design / full-stack
Role
Hotel procurement workflow covering bid collection, review scoring, contract review and archival audit, spanning the agent runtime, backend services and the React workbench.
Actions
  • Orchestrated a supervisor agent and multiple specialist sub-agents across technical, commercial and contract-compliance review.
  • Worked with domain professors to encode supply-chain, negotiation and legal-compliance methodology into each sub-agent's scoring rubric and expert-role persona.
  • Built Slock-style 'group-chat' collaboration: sub-agents join a shared channel as persistent members with identity and memory, exchange structured messages over shared context, with the supervisor handling turn-taking and @-mention handoffs.
  • Persisted the multi-agent discussion as structured messages and run events, enabling review summaries, online model tuning and run-cost tracing.
  • Added Cumora-style built-in agents (pre-defined personas/behaviors); since Slock and Cumora both run via a local npx server for data localization and this internal system has no such need, ran the agent runtime directly server-side — a single server handles all interaction, no per-client local node.
Impact
Moved procurement review into an auditable multi-agent group-chat decision loop where specialist agents converge in one room, with recoverable context across interruptions and reruns.

Etyon desktop AI workbench

· desktop AI workbench agent runtime / desktop
Role
Desktop AI workbench for local development and knowledge workflows, spanning the agent harness, chat runtime, project context and local CLI integration.
Actions
  • Built a harness layer across the AI SDK / Hono / Electron stack, separating model calls, context assembly, tool execution, approvals, sub-agent delegation and event logging into observable modules; wired in Skills, @ file mentions, a read-only plan mode and the Rust CLI, with multi-provider support (Anthropic / OpenAI / AI Gateway).
  • Deliberately narrowed the workspace tool surface to a minimal, auditable set (read / ls / grep / edit / write, writes approval-gated), backed by secret-path filtering and secret redaction before any tool data is persisted.
  • Recorded every run, tool call and approval to an append-only local event history (event sourcing) alongside the chat snapshot — run history and approval state survive a crash, pending approvals resume after a restart, and any run can be reconstructed in the run inspector as a tool-call and event timeline.
  • Shipped six built-in agent profiles (explore / coder / plan / review, etc.) plus read-only sub-agent delegation with hard limits — a concurrency cap, output clamps, no nesting; long-term memory is a resident per-project digest plus on-demand memory tools that keep per-turn context cost bounded.
  • Distilled a reusable Vite+ and Electron desktop scaffold along the way (main/renderer structure, build, packaging, CLI integration, i18n and project checks).
Impact
Moved an ordinary chat client into a recoverable local agent harness — every run leaves a full trace in the run inspector, paused approvals survive restarts, and one engine serves both the desktop app and the Rust CLI.

Internal knowledge agentic system

· agentic product direction lead
Role
Agentic chat and spaced-review system for knowledge management, spanning the web app, iOS surface, retrieval backbone, graph context and rich-text editing.
Actions
  • Orchestrated Mastra agent tool calls with approval gates on data-mutating note ops; read-only tools ungated.
  • Rebuilt hybrid retrieval across LLM query rewriting, FTS, title match, tag inverted index, pgvector embeddings and an LLM reranker.
  • Turned /ref and tags into xyflow-backed graph context for multi-hop reasoning.
  • Added a Japanese learning track to the review system: a reading-analysis view (furigana, POS tagging, bunsetsu segmentation) and JLPT N1–N5 exam practice.
Impact
Covered 5 core product loops and made retrieval thresholds, token / character budgets and wasTruncated markers observable enough to control BYOK context cost.

Cursor hackathon · AI coding agent

· AI coding harness architecture owner
Role
Claude Code-style programming assistant that organizes sessions, tools, plan/exec, build work and context compaction into a runnable coding harness.
Actions
  • Owned plugin selection, coding standards and the overall architecture.
  • Built the session lifecycle and a streaming TUI/CLI interaction layer.
  • Implemented file ops, code search, script execution, LSP checks and web search.
Impact
Delivered 20+ tool capabilities plus Plan / Exec, Compaction and Build sub-agents, allowing complex coding tasks to be split, executed and compacted.

Encrypted IM client

· cross-platform client core engineer
Role
Matrix-based Android, Electron and web client work covering rich text, deeplinks, navigation structure, deployment and Rust SDK maintenance.
Actions
  • Customized Matrix rich-text parsing and invite-link navigation.
  • Rebuilt the app-level deeplink dispatcher and decoupled flows through Appyx.
  • Maintained the Rust SDK and Debug / Release deployment paths.
Impact
With quality gates around AI-assisted coding, roughly 80% of business and boilerplate code was AI-generated, lifting delivery throughput about 3x.

Quantitative trading desktop

· live-trading app technical rebuild
Role
Desktop client for live quantitative trading, covering build architecture, strategy editing, data visualization, automated release and error-context governance.
Actions
  • Replaced electron-forge + webpack with electron-builder + Rspack.
  • Integrated TradingView, Monaco Editor and a GitHub Actions release flow.
  • Moved performance-critical JS modules into a Rust native addon (loadable in Electron) via napi-rs.
Impact
Cut compile/package time from 18s+ to under 4s and established multi-target distribution across GitHub Release, S3 and self-hosted endpoints.

Nebula low-code platform

· low-code SaaS frontend engineer
Role
Frontend and mobile migration work for an enterprise low-code platform, spanning workflow approval, maps, material persistence, runtime theming and legacy reuse.
Actions
  • Authored team collaboration standards and shipped an internal CLI.
  • Moved large material data from Web Storage to IndexedDB and migrated CSS Modules to CSS-in-JS.
  • Reused legacy flows through Module Federation to shorten the migration cycle.
Impact
Supported private deployment delivery that contributed to a 1M RMB-tier project while reducing merge conflict, review and migration overhead.

Education

Sichuan University Jinjiang College B.Eng. · Software Engineering