Hi, I'm Wang Hao!

Project Description

An enterprise Agent Skill factory for internal business scenarios. The platform enables business users to generate ERP / business Skills through natural language, then evaluate, publish, monitor, and optimize them through a governed lifecycle. It focuses on three key problems: uncontrolled Skill generation quality, possible fake tool-call results from Agents, and lack of observability and self-healing after release.

Responsibilities

• Contributed to the lifecycle pipeline covering Intent, Generate, Eval-Cases, Eval-Run, Publish, Monitor, Report, and Optimize.
• Designed the Stage 3 caseInputs + real spawn evaluation mechanism, where the system executes baicao api call and records stdout, exitCode, and duration.
• Built evaluation categories such as normal, edge, error, security, off_topic, and multi_turn, with Critical gates, Pass@K, Baseline Regression, and rubric scoring.
• Introduced LLM-as-Judge for structured assessment of open-ended outputs while keeping deterministic assertions for stable checks.
• Implemented Fix-Suggester menu-style self-healing and Telemetry instrumentation covering success rate, latency, tokens, toolchain, cohort, and content fingerprints.

Project Description

A unified CLI for the ERP capability gateway, serving developers, operations engineers, and Agent hosts. It provides standard entry points for authentication, capability discovery, parameter templates, API calls, KB retrieval, and diagnostics. Instead of asking Agents to guess APIs, parameters, or invocation methods, the CLI exposes auditable, testable, and recoverable access to ERP capabilities.

Responsibilities

• Designed Agent-friendly command surfaces covering api list/read/call/call-template, catalog search, invoke, and kb search/resolve/doctor.
• Designed structured outputs, parameter templates, and actionable error messages so Agents can recover from missing parameters, auth failures, and unavailable capabilities.
• Contributed to the KB v4 six-layer architecture: scenario -> workflow -> bridge -> clarify -> capability -> response, using SQLite FTS5 + BM25 + routing enhancement instead of raw vector RAG.
• Promoted Local-first design, Graphify 12 tables, manual / derived / view layer isolation, SHA-256 content addressing, commit log, manifest_root_hash, and C1-C6 recovery contracts.
• Built 7 test suites with 234 cases covering KB retrieval, capability discovery, parameter construction, and real API invocation; the KB retrieval suite passed 153 / 153 cases.

Project Description

I worked on critical modules in the self-operated pharmaceutical supply-chain ERP system, including finance, orders, work orders, customers, and marketing. As the team moved toward frontend-backend integrated delivery, the original ERP had 600+ pages concentrated in a single repository, with heavy coupling across shared dependencies and global state. This created high upgrade risk, expensive regression testing, slow rollback, and poor AI generation efficiency due to oversized repository context. I led the transformation into an AI-friendly architecture organized by business modules, independently deployable submodules, configurable routing, and fast rollback.

Responsibilities

• Defined module boundaries across finance, orders, work orders, customers, and marketing, and designed an "ERP shell + independent submodules" architecture.
• Built automated splitting scripts by page, module, and multi-module scope, replacing manual migration and reducing human error across 600+ pages.
• Split ERP modules into independently deployable Monorepo sub-repositories, embedded through iframe integration while sharing a unified entry component.
• Designed coexistence of old and new routes, allowing each module to switch between the new submodule and the original page component through online route configuration.
• Extracted shared business components and utility methods into npm packages, while moving route and tab configuration online to avoid repeated maintenance in submodules.
• Used iframe communication to synchronize global state, including tab state, page cache, refresh behavior, route parameters, and page switching after data updates.

Project Description

An enterprise Excel processing Agent that allows users to describe analysis, cleaning, merging, and multi-file comparison tasks in natural language. The Agent generates and executes pandas code, streams progress through WebSocket, and returns S3 files or text results. The goal is to let non-technical users complete complex spreadsheet work without writing scripts.

Responsibilities

• Designed a two-layer LLM architecture: the outer AgentExecutor decides whether to call the Excel Tool, while the inner ExcelProcessor generates and executes pandas code.
• Used Structured Chat + JSON blob to support GPT-4.1, GPT-4.1-mini, and Doubao-deepseek-v3, with handle_parsing_errors=True for format-error self-healing.
• Implemented AgentProgressTracker + WebSocket protocol, converting intermediate_steps into frontend-consumable fields such as percentage, step, iteration, and chat_response.
• Built SafeOS + thread-pool sandboxing to limit file and system access while supporting Excel analysis, multi-file comparison, PDF table processing, and chart generation.