Repetitive knowledge work
Hours spent copying between systems, summarising, classifying and drafting the same kinds of documents every week.
Independent AI studio · Remote worldwide
AI that does real work inside your business — scoped, measured, and handed off.
We help small and mid-sized teams turn large language models into dependable tools: agents that complete multi-step workflows, assistants that answer from your own documents, and evaluation that tells you whether the output is actually good.
- English & 中文
- Based in UTC+8, working remotely
- NDA available on request
- 1ReadTicket + customer history
- 2RetrievePolicy docs & past answers
- 3PlanDecide: answer, refund, escalate
- 4ActDraft reply · update CRM
- 5CheckRubric score before sending
Problems we solve
Where teams usually get stuck with AI
Most teams don't need more AI tools. They need one workflow that reliably works and a way to prove it.
Docs nobody can search
Policies, specs and past answers spread across drives, chats and wikis — so people ask colleagues instead.
Pilots that never ship
A promising demo that never reaches production because nobody defined success, edge cases or ownership.
No way to judge output
You can't tell whether a model or prompt change made answers better or worse — so every change is a guess.
Support load keeps growing
Repetitive customer questions eat team time while response quality depends on who happens to be on shift.
Content ops bottlenecks
Producing, formatting, translating and reviewing content at volume without losing brand voice or accuracy.
Services
What we build
Each service is scoped around one concrete workflow and ends with something you can run, measure and own.
AI agent & workflow automation
Multi-step agents that plan, use your tools and APIs, remember context, and hand off to a human when they should.
- Who it's for
- Ops, support and back-office teams with a repeatable multi-step process.
- What you get
- Workflow map with automation boundaries
- Working agent (LLM + planning + memory + tool calls)
- Guardrails, logging and human-in-the-loop steps
- Evaluation set and runbook
Typical timeline3–6 weeks
RAG knowledge assistant
An assistant that answers from your own documents and data, cites its sources and respects access rules.
- Who it's for
- Teams with large internal docs, product catalogs or policy libraries.
- What you get
- Ingestion and chunking pipeline for your sources
- Retrieval tuning and answer-with-citations
- Access-aware content and refresh process
- Accuracy report on real questions
Typical timeline3–5 weeks
LLM evaluation & quality harness
A repeatable way to measure model output quality, so prompt and model changes are decisions, not guesses.
- Who it's for
- Teams already running LLM features or choosing between models.
- What you get
- Task-specific rubric and scoring guide
- Benchmark set built from your real cases
- Side-by-side (SBS) comparison runs
- Regression checks you can rerun anytime
Typical timeline2–4 weeks
AI product strategy & rapid prototype
From a fuzzy idea to a clickable, working prototype that real users can try.
- Who it's for
- Founders and product teams validating an AI feature or product.
- What you get
- Use-case prioritisation and feasibility review
- Product requirements (PRD) and success metrics
- Working prototype on frontier models
- User test notes and next-step plan
Typical timeline1–4 weeks
AI content operations
Prompt systems and review workflows that produce, format and check content at volume.
- Who it's for
- Marketing, e-commerce and media teams with high content throughput.
- What you get
- Style and format specification
- Prompt templates and generation pipeline
- Automated format/quality checks
- Human review workflow and QA sampling
Typical timeline2–4 weeks
ChatGPT Business / Enterprise rollout & team enablement
Get your team actually using AI day to day — set up properly, safely, and with playbooks for real tasks.
- Who it's for
- Companies adopting ChatGPT Business or Enterprise (or similar) for a team.
- What you get
- Workspace setup, roles and data settings
- Custom GPTs for your recurring tasks
- Prompt playbooks per role
- Live training sessions and usage review
Typical timeline1–3 weeks
Engagement models
Start small, expand when it works
Clear scope and a fixed set of deliverables for every phase. Pricing on request — we quote after a short scoping call.
01
Discovery Sprint
1–2 weeks
Pick the right workflow and define success before building anything.
- Stakeholder interviews & process mapping
- Use-case shortlist with effort / impact
- Success metrics and data readiness check
- Written proposal for the build
Pricing on request
02
Prototype Build
2–4 weeks
A working version on your real data that users can try.
- Working prototype (agent, assistant or tool)
- First evaluation set and baseline scores
- Demo session with your team
- Gap list for production
Pricing on request
03
Production & Evaluation
3–6 weeks
Harden the prototype and prove quality before rollout.
- Integration with your systems
- Guardrails, logging and monitoring
- Full evaluation harness and report
- Deployment in your own accounts
Pricing on request
04
Ongoing Retainer
Monthly
Keep quality high as models, data and needs change.
- Regular evaluation runs and tuning
- New workflows and improvements
- Model / vendor updates
- Priority support
Pricing on request
How we work
Five steps, each with something you keep
- 01
Scope
Pick one high-value workflow, map it, and agree what 'good' means in measurable terms.
You getScope doc, success metrics, data checklist
- 02
Prototype
Build a working version quickly on frontier models and your real data.
You getRunning prototype and demo
- 03
Evaluate
Score it against a rubric and benchmark set; iterate on prompts, retrieval and agent logic.
You getEvaluation report with before/after scores
- 04
Ship
Integrate, add guardrails and monitoring, and deploy inside your accounts.
You getProduction deployment and runbook
- 05
Hand off
Train your team and transfer everything needed to run and extend it.
You getDocs, eval sets, code and training session
Example use cases
What this can look like in your industry
Illustrative examples of what we can build — not past client projects.
E-commerce & local services
- Product Q&A assistant grounded in catalog and policies
- Review and complaint triage with suggested replies
- Bulk listing copy generation with format checks
Automotive & in-car assistants
- Owner-manual and service-knowledge assistant
- Evaluation harness for voice-assistant answers
- Multi-turn dialogue testing for edge cases
Content & media
- Editorial workflow: brief → draft → fact/format check
- Multimodal output review rubrics (text, image, video)
- Translation and localisation pipeline with QA
Finance operations
- Invoice and expense document extraction
- Report drafting agent with source-linked figures
- Policy Q&A for internal finance questions
Principles
How we handle your data and your project
Your data stays yours
Work runs in your own cloud and AI vendor accounts wherever possible. We sign an NDA on request and only access what the project needs.
Measured, not guessed
Every build ships with an evaluation set and scores, so quality is visible and changes are comparable.
Documented handoff
Code, prompts, evaluation sets and runbooks are delivered so your team can operate it without us.
No lock-in
Standard tools and vendor-neutral design. You own the deliverables and can switch models or providers.
Founder
Xiangye (Xavier) Zhao
Founder · AI product & evaluation
Focus areas
- LLM agents
- Multi-agent systems
- RAG
- Prompt engineering
- Model evaluation
- Data & business analysis
- AI-assisted full-stack prototyping
Xavier works on AI products day to day — designing agent workflows and, above all, measuring whether they actually work. YeShenworld brings that practice to teams that don't have an in-house AI group.
- AI Product Manager — Meituan2025 – present
Evaluation and optimisation of a multi-agent system that generates business-analysis reports: instruction-following evaluation, prompt tuning, A/B testing and a reward benchmark for model quality.
- AI Trainer — Baidu2025
Prompt-engineered formatting correction for Markdown/LaTeX output; built three-axis evaluation rubrics (instruction following, physical realism, aesthetics) and side-by-side tests for text-to-video models.
- Education
B.Eng., Internet of Things Engineering — Harbin Normal University
FAQ
Common questions
Which languages do you work in?
English and Simplified Chinese — for meetings, documentation and the AI systems themselves (bilingual assistants are common).
Where are you based, and do you work remotely?
We're based in UTC+8 (Asia/Shanghai) and work remotely with teams worldwide. We schedule calls in overlapping hours, and most work happens asynchronously with written updates.
How do you handle data security?
Wherever possible, systems are built and run in your own cloud and AI vendor accounts, with your access controls. We sign an NDA on request, use the minimum data needed, and don't reuse your data for other work.
Which models and vendors do you use?
We're vendor-neutral. We typically work with OpenAI and other frontier and open-weight models, and choose based on your quality, cost, latency and data-residency needs — measured with an evaluation set, not assumed.
What size of project do you take on?
Focused projects: one workflow or product feature at a time, usually 1–6 weeks per phase. Larger programmes are broken into phases with a review between each.
How do we get started?
Send a short note through the inquiry form or email. We'll set up a 30-minute call, then send a written scope and quote. Most projects start with a Discovery Sprint.
Who owns the code and IP?
You do. On payment, deliverables — code, prompts, evaluation sets and documentation — belong to you, as set out in the project agreement.
What happens after delivery?
Every project ends with a handoff and training session. A short support window follows; ongoing tuning and new workflows are available on a monthly retainer.
Do you only build with ChatGPT?
No. We can help roll out ChatGPT Business or Enterprise for your team, but custom agents and assistants can use whichever model fits the task best.
Can you work with our existing engineers?
Yes. We often build the AI layer and evaluation while your team owns integration, and we document everything so they can take over.
Contact
Tell us about your workflow
A few lines is enough. We'll reply by email to set up a short call.
Prefer to write directly?
- Emailxavier@zxyshen.com
- ResponseBy email, usually with a call slot
- HoursUTC+8, with overlap for Europe & Americas
- LanguagesEnglish · 中文