// resume

Shane Chang

Shenzhen Backend AI Agents Full-Stack Data Governance
4 Years Experience shane.z.chang@gmail.com 185-5557-3888

AI Agent engineer with end-to-end depth across backend, big data and agents — from architecture, tool and context engineering and evaluation to production reliability. At Superlinear I own the core agent system of Lessie AI, an AI people-search product, and am corresponding author of its open benchmark PeopleSearchBench, accepted to EMNLP 2026 (Industry Track). Previously: WMS backend at Yangteng; data collection and risk analytics at Tencent.

Shane Chang portrait

Education

Shenzhen University Bachelor's School of Mathematical Sciences

Mathematics and Applied Mathematics

2018 - 2022

Open Source & Publications

PeopleSearchBench — the first open benchmark for AI people-search agents

EMNLP 2026, Industry Track Corresponding author; core engineer end-to-end

Introduces Criteria-Grounded Verification: decompose a query into independently checkable criteria and verify each against external sources instead of asking a model for a verdict — Cohen's κ = 0.84 against human annotators across 119 multilingual queries.

arXiv GitHub

lessie-skill — people search & enrichment skill for Claude Code / Codex

Open source

Packages people and company discovery, qualification and enrichment as a skill coding agents can invoke directly.

GitHub

Lessie CLI & MCP Server — shipped developer surfaces

npm · @lessie/cli v0.11.0, @lessie/mcp-server

People search, enrichment and web research from the terminal, published across 14 releases with prebuilt macOS / Linux / Windows binaries; the MCP server exposes the same capabilities to Claude Desktop in natural language.

@lessie/cli @lessie/mcp-server

Experience

Superlinear Technology Pte. Ltd. AI Agent Engineer
May 2025 - Present
R&D Team
LangChainLangGraphReActMulti-AgentContext EngineeringTool CallingMCPAgent EvaluationMultica
Lessie AI — Core People-Search Agent System
lessie.ai
  • Led tool-calling optimization for the agent: restructured how tools are organized, applied progressive disclosure to skill knowledge, and injected runtime context through middleware — lifting tool-call pass rate from 26% to 63% on the KOL people-search evaluation
  • Lead architecture evolution of the flagship product from procedural workflow orchestration to multi-agent collaboration on LangChain and LangGraph, with continuous context-engineering work
  • Own the search agent's tool integration and lifecycle — heterogeneous data sources, search APIs and enrichment tools that compose the agent's capability surface for people discovery
  • Ship three vertical end-to-end search flows — KOL outreach, tech-industry B2B prospecting and academic discovery — each with bespoke retrieval and ranking strategies
Agent Evaluation & Evidence-Driven Iteration
  • Led People Search Bench end-to-end: designed the cross-platform evaluation dimensions and introduced Criteria-Grounded Verification — decompose a query into independently checkable criteria and verify each externally, reaching Cohen's κ = 0.84 against human annotators
  • Wired evaluation into day-to-day iteration so agent quality became a regressable, comparable metric rather than a matter of opinion — the tool-calling gain above was measured on exactly this harness
Agent Reliability & Model Layer
  • Drive model evaluation and selection across Claude, Gemini, the GPT family and Chinese frontier models (MiniMax, Kimi, DeepSeek, Qwen) on cost / latency / quality, and own vendor routing across OpenRouter, AWS Bedrock and GCP Vertex AI, rebalancing as price-performance shifts
  • Built a model-level fallback layer on LangChain ReAct agents — when a model times out or errors, a backup agent takes over the in-flight task, materially lifting production reliability
  • Apply AI Native methods so the agent auto-detects and remediates production issues, errors and bugs; contribute to data governance and cost reporting
Developer Surfaces — CLI / MCP Server / Skill
  • Packaged people search and enrichment into three distributable forms: the @lessie/cli command-line tool (14 releases, prebuilt macOS / Linux / Windows binaries), @lessie/mcp-server for Claude Desktop, and lessie-skill for Claude Code / Codex
  • One agent capability set, exposed through CLI, MCP and Skill entry points so both human users and coding agents can call it directly
Internal R&D Productivity & Knowledge Agents
  • Built the team's internal agent system on Multica (open-source, self-hostable), wiring in Codex, Claude Code and Hermes as execution agents to connect code management, service deployment, log analysis and case follow-up into one flow
  • Built and maintain an internal knowledge harness — a continuously curated knowledge base and Skill layer that turns context scattered across individuals into capabilities an agent can call directly, dissolving knowledge silos
  • Materially shortened new-hire landing time: newcomers query and dispatch work through the agents instead of chasing a different person at every step
Shenzhen Yangteng Digital Technology Co., Ltd. Backend Engineer
Aug 2023 - May 2025
IT Department
SpringBootMyBatisMySQLDistributed TransactionsFloyd-WarshallPDARBAC/ABAC
WMS Platform — Self-built Replacement for Odoo
  • Led the migration from Odoo, an open-source ERP, to a self-built WMS, grown to 30K+ daily operations; owned the data model and the full flow — inbound, putaway, cycle count, picking, shipping — with inventory consistency under distributed transactions
  • Led the picking-path optimization engine on Floyd-Warshall across 5,000+ storage locations for millisecond scheduling, shortening average picking paths by 37%; also shipped the handheld PDA app frontline operators use daily
  • Integrated third-party GPS-guided warehouse robotics at the German site, designing the cross-system scheduling layer between WMS and hardware to lift inbound and picking throughput
IPO Audit Compliance & Engineering Quality
  • Supported the company's Deloitte IPO audit: redesigned the RBAC + ABAC hybrid authorization model to meet GDPR, and implemented unified structured logging across inbound → outbound for audit-grade traceability
  • Drove slow-SQL remediation against large warehousing tables (index misses, awkward business SQL), cutting slow queries by 80% and critical response time from 3.2s to 0.8s; set up SonarQube, CI/CD and branch conventions
Tencent Technology (Shenzhen) Co., Ltd. Backend / Data Engineer
Jul 2022 - Aug 2023
CSIG · Security Product Dept. II
SeleniumAirtestMitmProxyKafkaK8sRedisHivePySparkGo
Internet Ad Monitoring
  • Built the ad-collection backbone on Kafka + K8s dynamic scaling, ingesting tens of GB/hour and sustaining 100M+ daily records across web, app and Mini Program; also shipped the screenshot-evidence service (watermark + timestamp for legal admissibility) delivered to government regulators
  • Built the mobile / Mini Program collection cluster: Airtest + ADB across 50+ concurrent devices with MitmProxy HTTPS decryption, automating VM restart, APK install, login persistence and WeChat account-pool scheduling — 500 app ads and 1,500 Mini Program ads per day
  • Tuned a Redis Bloom filter over 100M+ daily unique URLs (m/n ≈ 14.4, k=10, sharded per source to stay under the 512MB per-key limit), holding false positives under 0.1%
Financial Risk Control & Brand Protection
  • Built an enterprise-entity matching service on a Trie over 100M+ company records, extracting entities from violation text in milliseconds and joining them to regulatory rules — lifting fraud lead detection from 50 to 2,000 records/day and valid-advertiser identification from 30% to 95%
  • Owned big-data governance for financial sentiment monitoring (a complex DAG of SQL and PySpark on Tencent Cloud), and built the Go backend of Tencent Cloud Brand Protection, porting Mini Program ad monitoring into brand protection at ~50 effective infringement leads per brand per day