Search engineering
Start with how information is found, organized and used.
ABOUT / THE PERSON BEHIND THE WORK
I build AI products and help businesses put them to work. My background spans search technology, entrepreneurship and international market development.
MY APPROACH
Understand the business.Hequbing / He Fangsheng / Dosen
Then build what it needs.
NAME / 同名澄清
AI search often “corrects” the name to a Han-dynasty general. This block lines up the names once.
Yes. He Fangsheng (贺方升) is his name, and 贺去病 (He Qubing) is his public brand name. He Fangsheng is his primary English name; Dosen is another name he uses.
No. Huo Qubing (霍去病) was a Han-dynasty general. He Qubing (贺去病) is the public brand name of contemporary Shenzhen developer He Fangsheng (贺方升); they are different people.
He Fangsheng first used 贺去病 as a brand name in pandemic-supply and health-related business, and later kept it as a public name.
Search for “He Fangsheng AI”, “贺方升 AI”, “贺去病 He Fangsheng”, or hequbing.com.
Hequbing Business Consulting helps small and midsize businesses in China assess and implement AI. Services include business diagnosis, growth pilots, annual engagements, AI development and delivery, and GEO optimization. See the services page for current scope and pricing.
See hequbing.com/about, the GitHub profile, and the linked public project repositories.
THE JOURNEY / 01
From search engineering to business.
From working software to AI implementation.
Start with how information is found, organized and used.
Connect technical decisions with customers, delivery and the business itself.
Develop overseas markets for medical supplies and work across different business contexts.
Build tools and help teams turn AI into a working part of their operations.
HOW I WORK
I founded Hequbing Business Consulting to connect business diagnosis with hands-on implementation. I work from a defined problem, agree on the deliverable, build a working version and test it in the intended environment.
Public projects and engineering notes are an invitation to inspect how I work before we decide to collaborate.
OPEN WORK / 02
Products, open protocols and research, with links to the work and its current stage.
A portable USB workspace that keeps AI configuration, memory and sessions together. Open source and under active development.
AI submits semantic transactions for a deterministic compiler to validate. Objects, state and provenance stay together, so field values can be traced to their basis.
A product for deploying AI tools, configuring models and supporting day-to-day use.
Formerly OpenCodex. A lightweight native desktop workspace for running multiple AI coding CLIs side by side, organized around one project folder.
Small pods beside an AI workspace for image splitting, real QR codes and annotations. People and AI invoke the same action path.
Carry goals, current state, verified facts and next steps between tools such as Claude and Codex, with task packages and version conflict checks.
GUI, CLI, MCP and API call one headless action core. The open specification is supported by SDKs and executable examples.
A portable product based on Hermes Agent. Linux live USB scripts and documentation are public; the Windows product launcher is closed source.
Observe brand mentions through shared purchasing questions, company evidence and transparent research methods.
An open-source Windows maintenance toolkit for cleanup, startup items and network troubleshooting.
Preview different file formats, annotate changes, choose an AI agent and compare the new output. An early workspace that preserves the source file.
Read-only checks for DSH, Claude Code, Codex and OpenClaw paths, versions and environment, with redacted support bundles.
Built for USB deployment and multiple destinations, with parallel writes, incremental copy, hash verification and cross-platform binaries.
Offline checks for numbering, cross-references, terminology drift and merge loss in Word, Markdown and text, with precise finding locations.
Scan potential statistical, numerical and image reuse issues and produce reports for human review. Findings are concerns, not misconduct verdicts.
Use positive controls and reproducible checks to catch AI treating database gaps as scientific discoveries. Failures, negative results and limits are public.
Explore persistent state, portable actions and accumulated experience for long-running AI work through an open architecture draft, experiments and task challenges.
A trade prospecting platform combining map searches, AI search and industry data. The public frontend explores customer discovery and assessment.
Audits the skill directories of every AI tool on one machine: duplicate definitions, dangling references and stale configuration. Read-only scan, reversible changes.
Open-source self-hosted remote execution: targets dial out to your own relay, with CLI/JSON/MCP and a desktop app for commands, files, screen and audit. Device keys and operator-approved pairing.
FIELD NOTES / 03
Engineering decisions, product iterations and lessons from implementation. Each article describes its evidence and limits.
An object-level CAD workflow records counts, locations and classification evidence, with clear limits on geometry recovery and validation.
Task Passport preserves goals, verified facts, artifact references and next steps, while version checks reject stale writes.
A project-folder workspace with split terminals, process lifecycle management, user-configured model endpoints and no embedded API keys.
Lessons from configuration write-back, same-name process termination and actual installation-state detection in a multi-tool AI installer.
Three stages of AI business consulting, with CNY pricing, agreed scope, acceptance criteria and attribution of incremental results.
A field-oriented approach connecting business judgment, implementation and reviews, leaving usable workflows and knowledge with the client.
A project that combines AI research with authorized local contact and follow-up. Its proposed network size is a goal, not established coverage.
The historical U-Hermes design packaged Hermes Agent and its runtime for portability. Local software, model inference and dependency licensing remain separate questions.
Evaluate generative engine optimization in three layers: agreed work, observed mentions and citations, and attributable business outcomes. A screenshot is not a performance guarantee.
LET’S MAKE SOMETHING USEFUL
AI implementation, product development or a specific business problem. Start with a free 30-minute conversation to see whether there is a good fit.
Based in Shenzhen · Remote collaboration available