AI Full-Stack Developer focused on building reliable, observable AI products that solve real workflow problems.

Currently an AI Full-Stack Developer Intern at ByteDance. Previously a Frontend Developer Intern at Kingsoft.

I care about the engineering around the model: retrieval quality, tool boundaries, human review, evaluation, observability, and product experiences people can actually use.

A local-first developer tool that finds visual inconsistencies in rendered web interfaces and turns them into explainable offline reports.

What it demonstrates:

• Chromium-based collection of computed styles and element geometry
• Conservative peer grouping before evaluating spacing, color, typography, radius, and alignment
• Self-contained offline reports with overlays, filters, search, keyboard navigation, and evidence
• Privacy-first execution with no telemetry and isolated browser state
• 13/15 labeled inconsistencies detected, 0 medium/high findings on the clean fixture, and a verified 1,001-element performance gate

Source · Release · Architecture · Release audit

An AI customer-support and ticket collaboration platform for SaaS teams. It turns scattered knowledge, repeated status checks, and risky AI replies into one traceable workflow.

flowchart LR
    Q["Customer question"] --> R["Hybrid RAG"]
    Q --> M["Read-only MCP tools"]
    R --> A["Eino Agent"]
    M --> A
    A --> G["Evidence & risk guard"]
    G -->|Safe| D["Cited response"]
    G -->|High risk / weak evidence| H["Human review"]
    H --> D

What it demonstrates:

• Hybrid PostgreSQL full-text + pgvector retrieval with traceable citations
• Go/Eino Agent runtime with three schema-validated, read-only MCP tools
• Human-in-the-loop review for high-risk or weak-evidence drafts
• Persisted Agent Runs, replay links, streaming events, and secret redaction
• Next.js customer portal, support workspace, knowledge operations, and evaluation UI
• Docker one-command demo with no model key required
• 24/24 deterministic orchestration and safety contract cases, with dataset hash and source commit recorded

Source · Architecture · Evaluation result · 3-minute demo

A local-first Codex skill for discovering Mainland China job openings and tracking applications in a private, editable dashboard.

What it demonstrates:

• Resume-guided job-search profiles without persisting personal contact details or full resume text
• Evidence-based discovery across employer career sites and public Mainland China recruitment sources
• Explainable role scoring, company-level recommendations, direct-link classification, and official application-limit tracking
• Token-protected localhost dashboard with auditable application-status history and privacy-filtered exports
• Python standard-library implementation backed by 103 automated tests and GitHub Actions CI

Source · Skill · Tests

Open Source Contributor to Alibaba & Tencent Projects — 4 pull requests submitted, with 3 merged and 1 under review.

• Tencent / WeKnora — Added native XMind document parsing and knowledge-base ingestion (#2713, merged)
• Alibaba / ANOLISA — Improved stale repository-source diagnostics and recovery guidance (#2650, merged)
• Alibaba / OpenCodeReview — Documented missing CLI review flags and usage (#900, merged)
• Tencent / WeKnora — Hardened LLM rerank fallback behavior for reasoning models (#2711, under review)

Go · Eino · MCP · PostgreSQL/pgvector · React · Next.js · TypeScript · Docker

Useful before impressive.
Observable before autonomous.
Reproducible before claimed.

I'm currently looking for opportunities in AI application development and AI full-stack development.