Profile
AI Platform Architect at Career International, with 14+ years in full-stack development and architecture. I lead the architecture and production launch of enterprise AI products, and independently build and continuously iterate aibuddy — a production-grade, general-purpose Agent platform.
Commercial Impact Led architecture and core delivery for 3 commercial AI products: YiMai (JD-rewrite suggestions ~46% adoption) · MS intelligent search · Mira (research, document handling and executive-search sourcing; 50–70% efficiency gains)
Core Engineering Context engineering sets the ceiling for long-horizon Agents: multi-tier compaction + token-budget calibration + checkpoint recovery, keeping the Agent running stably within a bounded context window without losing task state
End-to-End Ownership Built and continuously iterate aibuddy from 0 to 1 — a production-grade general-purpose Agent platform; decoupled by capability into composable modules so one Agent core deploys to both server and desktop
AI Builder — Personal Project
aibuddy
From problem definition to continuous iteration
A production-grade, general-purpose Agent platform I develop and continuously iterate independently — supporting general Agent task execution, Agent Team collaboration, scheduled jobs and structured delivery; spanning the full lifecycle from product definition and architecture to core development and delivery.
Tech StackTypeScript · React · AI SDK · Hono · Drizzle · BullMQ · Electron
Runtime & IsolationPostgreSQL / Redis · SQLite · Node / Daytona / E2B Sandbox
Context Compaction EngineMulti-tier compaction with cross-turn snapshots, calibrating the token budget against actual usage; enables long-task context control and checkpoint recovery.
Multi-Model Adaptation & RoutingAdapts to each model's tool-calling differences, loads tools on demand with graceful fallback; keeps Agent capability consistent while cutting context overhead and improving cache reuse.
Agent Team DAG OrchestrationParallel multi-Agent orchestration over a dependency DAG; wakes agents on demand once dependencies are met, with wake-locks preventing duplicate execution.
Dual-End Zero-Drift ArchitectureAbstracts storage, queue, transport and Sandbox into replaceable interfaces, so the same Agent core deploys to both server and desktop — avoiding dual implementation and keeping behavior and task execution identical.
Engineering Practice Agentic Coding workflow: 15 workspace packages managed with pnpm + Turborepo; architectural context and delivery standards captured as versioned Skills; automated quality gates via TypeScript strict, Zod, lint and build.
Experience
Career InternationalHR SaaS · Recruitment Platform
Dec 2020 – Present · Suzhou
NowAI Platform Architect Jun 2023 – Present
Driving the productization of AI across vertical recruitment and talent-pool scenarios: delivered YiMai's generative-AI features in 2023, rebuilt MS intelligent search in 2024, and since 2025 have led the Mira Agent from product definition to launch; responsible for architecture and core delivery, validated through production usage data and offline evaluation.
- 2023 YiMai · Generative-AI Productization Brought résumé-parsing optimization and JD rewriting into real user flows, reaching ~46% adoption of JD-rewrite suggestions; delivered structured I/O, multi-client streaming interaction and business-system integration; owned the AI feature architecture and core development
- 2024 MS · Talent & Résumé Intelligent Search Upgraded the central talent pool from keyword search to hybrid retrieval — Embedding + BM25 + business-rule recall + rerank — extracting structured conditions such as role, skills, industry and experience; built a Python offline-evaluation harness with LLM-as-a-Judge to validate semantic gains in this keyword-heavy recruiting domain, lifting nDCG@20 from 0.76 to 0.80 and enabling fast iteration
- 2025–Present Mira · Agent Architecture & Business Adoption Led the recruiting Agent product from definition to launch and into real business use. Built three scenarios around headhunters' core work: Research assists consultants with information retrieval, analysis and delivery (~60% efficiency gain); daily document handling covers attendance, finance and similar tasks (~70%); Sourcing targets executive search, improving candidate search and screening efficiency (~50%). Captured recruiting judgment rules and delivery standards as reusable Skills; standardized capability interfaces, permission boundaries and error handling via a unified Tool contract so the Agent executes safely and reliably; controlled long-task context length through compaction, with human confirmation (HITL) and scenario-based permissions on critical actions
Senior Full-Stack Engineer · Tech Lead Dec 2020 – Jun 2023
Initially led the client-side (C-end) rebuild of YiMai, keeping core paths stable and supporting product iteration; later served as team lead, owning end-to-end delivery of ShuKe from initial scope definition to production launch and subsequent iteration.
- YiMai C-End System Rebuild The mini-program is the company's primary user entry (3–5K DAU); to support ongoing iteration, rebuilt the backend APIs and core domain models and led a Spring Cloud microservices migration — layered into adaptation, application, domain and infrastructure — forming a shared service foundation for both the mini-program and the app
- ShuKe 0→1 Business Incubation As team lead, drove ShuKe from product incubation to launch and operation, owning requirement assessment, solution decisions, core development, task allocation and delivery-cadence management
- Data & Recommendation: Integrated the vertical recruiting system with the company data platform, building data sync and processing pipelines on DataWorks; contributed to recruiting-domain modeling and unified metrics for roles, candidates, impressions, engagement and conversion; rebuilt recommendation and distribution logic on Elasticsearch, lifting core recommendation-distribution metrics by up to 25%+
- Engineering Efficiency & Stability: Helped build the CFD cross-functional collaboration mechanism and the LightMerge multi-branch integration approach, improving requirement flow, branch collaboration and incident response, and cutting system MTTR by ~20%
Huaer YingsaiCross-Border E-Commerce · ERP
Sep 2018 – Dec 2020 · Suzhou
Tech Lead · 2 yrs 3 mos
Built a 7-person engineering team from scratch and owned a cross-border e-commerce ERP from business analysis and architecture to launch; the system spanned procurement, inventory, orders, finance and multi-platform integration, with accountability for scope, technical architecture and delivery quality.
- Team & Delivery: Built a 7-person engineering team from scratch, breaking procurement, inventory, order and finance requirements into a first-release scope and iteration plan, driving the ERP to first launch and supporting daily operations
- Modular-Monolith Architecture: Given fast-shifting early business boundaries and team size, adopted a Spring Boot modular monolith to avoid premature microservice splitting; partitioned by business domain and governed the responsibilities and dependencies of 20+ core modules, balancing architectural evolution against maintenance cost
- Full-Chain Integration: Connected 5+ systems — Amazon, Walmart, Yum ERP, an in-house ERP and standalone sites — through an internal canonical data model and adaptation boundaries, unifying procurement, product, order and inventory data across the chain and turning manual reconciliation from routine work into exception handling
PatSnapSaaS · Patent / IP
Nov 2016 – Sep 2018 · Suzhou
Senior Backend Engineer · 1 yr 10 mos
Built configurable form, workflow and data-collaboration capabilities to meet enterprises' differing process and data-structure needs, reducing repeated customization cost in SaaS scenarios.
- Form & Workflow Configurability: Separated the stable form-runtime mechanism from enterprise-specific configuration, designing a configuration model spanning form definition, rendering, validation, BPM orchestration and data distribution; supported 50+ business forms and cut repeated customization for new forms and clients
- Event-Driven Data Collaboration: Designed the post-submission data routing and distribution pipeline, connecting BPM, the search engine and multiple business microservices per configuration; used Kafka for async decoupling, with retries and compensation to guarantee cross-service eventual consistency
Haibei Network TechnologyInternet · Mobile App
Jul 2014 – Nov 2016 · Suzhou
Full-Stack Engineer · 2 yrs 4 mos
Independently built the mobile-app backend on top of an existing web business, and continuously improved search, performance and automated delivery.
- App Backend Build: Independently designed and built the mobile-app backend on top of the existing web system, standing up authentication, content services, search and push notifications within 2 months to ensure an on-time launch
- Search & Performance Optimization: Introduced Elasticsearch for content search and Redis caching for high-frequency APIs, reducing hot-path latency and improving service stability
- Automated Delivery Pipeline: Introduced Docker containerization and Jenkins CI/CD, establishing a continuous delivery flow across commit, build and release
Ci123Internet · Parenting Community
Jan 2013 – Jul 2014 · Nanjing
Full-Stack Engineer · 1 yr 6 mos
Developed activity and interaction features for a parenting community, covering requirements, front- and back-end implementation, system integration and launch.
- Campaign Delivery: Independently delivered marketing campaign pages end to end — requirement breakdown, page implementation, interaction development, API integration and launch
- Community Interaction: Built commenting, likes and sharing, and integrated the commenting system with the forum team