Skip to content
View shanezchang's full-sized avatar

Block or report shanezchang

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
shanezchang/README.md

Shane Chang

AI agent engineer in Shenzhen. At Superlinear I work on the core agent system behind Lessie AI, an AI people-search product. Most of my recent work is on agent evaluation: checking whether an agent's output is actually correct, without leaning on another model's opinion.

Resume · 中文简历 · LinkedIn

Publication

PeopleSearchBench: Evaluating AI-Powered People Search Platforms with Criteria-Grounded Verification
EMNLP 2026 Industry Track · corresponding author · code

An open benchmark of 119 multilingual queries across recruiting, B2B prospecting, expert search and influencer discovery. Rather than asking an LLM to grade results, each query is broken into criteria that can be checked independently, and every returned person is verified against live web evidence. Agreement with human annotators: Cohen's κ = 0.84.

Open source

  • people-search-bench: the benchmark and evaluation harness from the paper
  • lessie-skill: people and company search and enrichment, packaged as a skill for Claude Code and Codex
  • @lessie/cli and @lessie/mcp-server: the same agent capabilities as a CLI and an MCP server
  • mcp-hub: my own MCP server hub, context and conventions for AI coding tools
  • Sideword: offline, keyboard-first English vocabulary practice in the terminal
  • Plotip: IP lookup on a map, plus browsing a region's IPv4/IPv6 ranges

What I work on

  • Tool calling and context engineering. Restructured tools, progressive disclosure of skills and middleware context injection took our people-search eval pass rate from 26% to 63%.
  • Agent evaluation. Criteria-grounded verification as an alternative to LLM-as-judge.
  • Multi-agent systems in production. LangChain / LangGraph orchestration, model routing and fallback.
  • Agents for internal R&D. Codex, Claude Code and Hermes wired into code, deploys, logs and support cases through Multica.

Before agents

Backend engineer at Yangteng, where I led the move from Odoo to an in-house warehouse system handling 30K+ operations a day that passed a Deloitte IPO audit. Before that, backend and data engineer at Tencent (CSIG), on large-scale ad-compliance monitoring built on Kafka and Kubernetes.

B.Sc. in Mathematics, Shenzhen University.

Pinned Loading

  1. LessieAI/people-search-bench LessieAI/people-search-bench Public

    The first open benchmark for evaluating AI-powered people search agents

    Python 154 36

  2. LessieAI/lessie-skill LessieAI/lessie-skill Public

    A Claude Code/Codex skill for searching, qualifying, and enriching people and companies. Supports candidate sourcing, B2B lead generation, contact enrichment (email, phone, LinkedIn), company resea…

    34 4

  3. sideword sideword Public

    Offline, keyboard-first English learning in your terminal

    Python

  4. plotip plotip Public

    Plotip / 落点 — open-source IP geolocation and regional IP range exploration. FastAPI, uv, TypeScript and MapLibre. No accounts or paid map APIs.

    TypeScript

  5. shanezchang.github.io shanezchang.github.io Public

    Personal site and bilingual resume (Astro, GitHub Pages)

    Astro

  6. francozanardi/html2pic francozanardi/html2pic Public

    Convert HTML + CSS to images in Python. Fast, without a browser.

    Python 33 3