Zhao Junquan

General Tool Platform (AiNative-Tools)

Core Project Iterative Dev AI-Native Studio LangChain-style Workflow DEMO

Project Background

Built for cross-border e-commerce content production, automated product modification, image processing, and batch delivery, AiNative-Tools integrates product collection, Prompt orchestration, AI copywriting, white-background images, scene visuals, package delivery, and runtime observability into a local Studio workspace. The system uses directory contracts, task queues, plugin manifests, and status streams to organize sourcing, listing, asset generation, and delivery into configurable, resumable, and traceable production workflows. Its engineering shape follows a LangChain Workflow around e-commerce operations: each product batch moves through stable input, Prompt orchestration, model calls, structured parsing, image tools, and delivery tools.

My Responsibilities

Owned the productization of cross-border e-commerce content workflows, turning collection, product rewriting, translation, image generation, and delivery into an end-to-end local Studio process. I led the scope definition from business workflow decomposition to tool-boundary design and was responsible for making repetitive operations configurable, traceable, recoverable, and scalable for later expansion.

Technical Stack

A FastAPI + TaskManager + Manifest 3-layer control plane, where tasks progress from ingestion to delivery in batches, with real-time WebSocket state sync.

The frontend Studio Shell manages task queues, batch operations, and tool entrypoints in vanilla JS; the backend handles native/cloud/AI task channels via FastAPI async routes.
The Manifest layer auto-scans local SUM tool scripts and registers image processing, packing, and enhancement capabilities as configurable tool forms.
Core production chain: product URL/Excel input → Scraper → PromptComposer → AIClient → JSON Parser → Painter/PackTool delivery.
AIClient normalizes Gemini / OpenAI-Compatible model protocol differences; Structured Output Parser constrains model output into JSON consumable by downstream image tools.
Prompt Registry supports market-language substitution, stage slicing, and Sandwich-format packaging so different product batches run stably through the same pipeline.
Key Pool, state snapshots, and batch directories together provide recovery boundaries for retries, missing images, re-translation, and re-packaging.

Key Milestones

The project upgrades scattered operation tools into a visual, queued, and recoverable Studio control plane. It supports a full loop from product URL / Excel input to image-text collection, AI analysis, white-background preprocessing, batch scene-image generation, and delivery package assembly, while improving long-batch stability through directory contracts, state snapshots, Key Pool scheduling, and log replay. Batch folders act as stable input Schema, the self-built TaskStateStore works as Checkpoint, SharedGroupManager provides Memory-like State, and image / packaging scripts join the workflow as Tool Layer components, preserving clear recovery paths for failed runs, reruns, image supplementation, translation supplementation, and repackaging.

FastAPI Python 3.12 Pydantic Native JS Studio Shell WebSocket State Stream Manifest Plugin System TaskManager Multi-Queue Gemini / OpenAI API Prompt Composer Key Pool Scheduling Selenium / Firefox Pillow / OpenCV XLSX / CSV Data Ingestion Batch Archive & Recovery Prompt Registry RunnableSequence Structured Output Parser State Store Memory-like State LangChain

View Dev Log

Lite Studio Client

Desktop Client Ongoing Maint

Project Background

As LLM tool-call complexity keeps growing, ad-hoc scripts can no longer sustain high-frequency, multi-turn, cross-toolchain AI workflows. Lite Studio replaces scattered utilities with a persistent, traceable, and extensible local AI workstation — consolidating conversation state, model configuration, context compression, and plugin orchestration into a long-term maintainable system.

My Responsibilities

Led the productization of Lite Studio from a collection of scattered AI utilities into a local desktop workstation. I defined the capability boundaries for conversation management, model configuration, context compression, plugin calls, and export flows, while establishing documentation, API contracts, and workflow baselines so future iterations could stay traceable and deliverable.

Technical Stack

pywebview hosts the local desktop runtime; Vue + SQLite + plugin system form the desktop AI workstation.

The pywebview desktop layer shares a process with the Python backend; the frontend calls LLMEngine / PluginHost via bridge.js, receiving streaming events and plugin state.
context_ledger appends conversation summaries as structured nodes instead of raw history stacking, providing a structured entry point for long-conversation compression and graph-based understanding.
bridge.js unifies streaming responses, chain-of-thought pushes, and plugin events into one standard event format for the frontend.
A 5-layer assembly pattern constrains module dependency direction: Storage → Gateway → Engine → PluginHost → Front Bridge, reducing implicit coupling over long-term iteration.
SQLite manages sessions, messages, presets, and context_ledger; all local data is self-contained without external services.

Key Milestones

The project documentation became the team's unified engineering reference, covering 80+ API contracts and 12 core workflows. context_ledger transforms long-conversation management from hard truncation to structured compression through incremental summary-node append. A standalone Web Demo page was built on top, recreating the full interaction — sidebar collapse, model switching, thought-process visualization, memory archiving, and multi-format export — in a vintage oatmeal design system, requiring no backend to showcase the complete product experience.

Python 3.8+ pywebview SQLite Tailwind / Vue Memory Context Ledger Graph

View Dev Log

OpenClaw Sentinel

Multi-Agent MIT License GitHub Repo

Agent Routing Guardrails + WebSearch Sub-Agent Orchestration

Project Background

In multi-tool Agent workflows, if the main Agent owns all search, research, and execution tools by default, high-cost APIs are easily overused and lightweight tasks get mixed with deep research in one context, causing cost waste, blurred boundaries, and security risk.

My Responsibilities

Independently owned the definition and delivery of OpenClaw Sentinel, abstracting search cost, tool permissions, and execution boundaries in multi-tool Agent workflows into a reusable routing-governance pattern. I led behavior-boundary design, tool-call policy, validation flow, and open-source release so one MCP integration became a portable Agent governance solution.

Technical Stack

A dual-layer Agent routing model: the main Agent handles lightweight search and execution; a web-researcher Sub-Agent handles deep research.

The main Agent is responsible for Q&A, code operations, task planning, and lightweight retrieval; the Sub-Agent only performs search/verification/reasoning/summarization and does not execute locally.
Presets declare allow/deny lists and tool-calling policies, preventing the main Agent from overusing high-cost research tools.
General questions default to DuckDuckGo lightweight search; only complex research, multi-source verification, architecture judgment, or explicit user requests route to Perplexity MCP.
The solution ships with verification scripts, install scripts, and bilingual README, forming a portable Agent tool-governance template.

Key Milestones

Delivers cost-tiered search vs deep research, clear execution-vs-research responsibility isolation, and strict security boundaries: no secrets in repo, no API key exposure, no real config uploads. The project ships with MIT license, bilingual README, presets, install script, and validation script for community evolution.

OpenClaw MCP Sub-Agent Agent Orchestration Perplexity API DuckDuckGo Agent Routing Tool Boundary

View Dev Log

MiMo Vision MCP

Vision MCP Plugin MIT License GitHub Repo

MCP Vision Bridge for Text-only LLM Agents

Project Background

Text-only LLMs (DeepSeek, Qwen, etc.) excel at reasoning and coding but lack native vision. This project wraps Xiaomi MiMo v2.5 as a standard MCP Server, giving text-only agents image understanding via the mimo_image_analyze tool — the main model reasons, MiMo perceives.

My Responsibilities

Independently owned the design and delivery of a vision-bridging solution for text-only Agents, turning image understanding from a single model call into an installable, configurable, and reusable MCP tool. I defined usage scenarios, call boundaries, secure configuration, and open-source delivery so OpenClaw, Claude Desktop, Cursor, and similar Agent clients could attach the capability reliably.

Technical Stack

Python MCP SDK wraps MiMo v2.5 as an MCP Server so text-only Agents can mount vision on demand.

The MCP Server handles protocol declaration, tool registration, and Stdio/SSE transport adaptation.
The MiMo API Adapter handles local-file/URL image encoding, API requests, and text result normalization.
Supports a dual-parameter interface: image_path (required) + prompt (optional), compatible with Claude Desktop, Cursor, Cline, OpenClaw, and other MCP clients.
API keys are kept out of the codebase and configured securely through environment variables.

Key Milestones

Implemented a complete MCP vision bridge in a single server.py, filling the vision gap in the Agent toolchain. It supports screenshot debugging, UI analysis, OCR extraction, chart interpretation, and document analysis. The design is loosely coupled, replaceable, and easy to orchestrate: the main model stays separate from vision logic, while vision is attached on demand. It has run stably in an OpenClaw Agent environment.

MCP MiMo v2.5 Vision AI Python Text-only LLM Image Analysis OCR Agent Tool

View Dev Log

QCode Toolkit (SVG QR Code)

Frontend-Only Tool Single File, No Build Required Try Online

Project Background

Because daily workflows often require batch QR-code generation and decoding, while many existing tools are ad-heavy, paid, or network-dependent, this project builds a no-install, backend-free, frontend-only QR-code utility.

My Responsibilities

Independently owned QCode Toolkit's product planning, interaction design, and release delivery, defining the tool boundaries around QR-code generation, styling, batch processing, and SVG download. The focus was closing the loop entirely in the browser without backend dependency while balancing usability, shareability, and long-term maintenance cost.

Technical Stack

Single-file HTML / Vanilla JS / CSS — generate, parse, style, and batch-process QR codes entirely in the browser, no backend needed.

The UI layer handles text input, drag-and-drop, style adjustments, and preview; the parsing layer integrates jsQR to recognize QR codes from images in the browser.
The rendering layer uses QRCode.js or custom SVG logic to produce PNG / SVG output.
The DOM Drag & Drop API supports multi-file batch processing; SVG export ensures clarity in print and secondary design scenarios.
The entire pipeline never touches a backend — no data upload risk, suitable for sharing as a standalone HTML file.

Key Milestones

Delivered a true browser-ready utility covering QR generation, styled redraw after decoding, and batch recognition while keeping data local, private, and portable across platforms.

Vanilla JS HTML5 / CSS3 Single File App Batch Processing SVG Export Drag & Drop API

Repo View Dev Log

QuanWen.art

Fullstack Site Content Hub AI-Native Bilingual

Project Background

QuanWen.art is a self-built personal digital content hub, designed as a long-term operable fullstack Web System. It integrates project showcases, photo essays, progress updates, AI chat, contact messages, and admin management into one architecture, giving the personal site content publishing, data maintenance, access protection, and scalable growth capabilities.

My Responsibilities

As the independent product and engineering owner of QuanWen.Art, I was responsible for the full planning and delivery of site positioning, content structure, public presentation, backend operation, and security maintenance. I integrated projects, photo essays, progress logs, AI chat, contact messages, and admin management into a long-running digital content hub while maintaining its updateability, accessibility, and security boundaries.

Technical Stack

FastAPI + Jinja2 partials + SQLAlchemy + MySQL, covering public SSR rendering, admin JSON APIs, AI chat, media management, and security hardening.

The frontend SSR layer uses 10 Jinja2 partial modules to organize homepage rendering, optimizing above-the-fold visibility via StreamingResponse and lazy-loading strategies.
The backend FastAPI routing and service layer handles content reads, guestbook, AI chat, media management, and admin operations.
The data layer uses 14 MySQL tables to manage entities for profiles, projects, photos, media, messages, access logs, and security logs.
Security layer: TrafficGuardMiddleware + CSRF + Session Token + IP rate limiting + Nginx rate limiting + fail2ban form a defense-in-depth system.
Tailwind CSS migrated from CDN runtime compilation to static compilation, eliminating all render-blocking resources; TTFB reduced from 262ms to ~20ms.

Key Milestones

Independently delivered a fullstack self-built Web System, upgrading QuanWen.Art from a static portfolio page into an operable digital content hub. The frontend uses StreamingResponse for chunked streaming rendering, with partial modules enabling earlier first-screen visibility and turning the homepage from a monolithic template into a maintainable modular structure. The backend connects public APIs, admin management, AI chat, media uploads, visit logging, and content services through FastAPI, forming a complete loop from presentation to operation. The data layer uses 14 MySQL tables to support profiles, projects, photos, messages, media, visits, and rate-limit records, with a DB-first / static-fallback strategy for better availability. The security layer covers login protection, CSRF, Session Token, IP rate limiting, traffic-tier control, abnormal request blocking, and Nginx/fail2ban coordination, giving the project the engineering completeness required for real production deployment.

AI-NATIVE Streaming FastAPI Jinja2 SQLAlchemy MySQL i18n Bilingual Content Hub Modular Partials Performance Optimization 2026

View Dev Log

Qwen3-8B Chinese Support Agent Finetuning

LLM Finetuning QLoRA Support Agent Kaggle T4

Project Background

This project targets the QuanWen.Art website support-agent scenario and explores parameter-efficient finetuning for an 8B open-source LLM under Kaggle's dual Tesla T4 resource constraints. Using Qwen/Qwen3-8B as the base model, it injects Chinese instruction data covering identity Q&A, technical explanations, product-copy generation, and support replies, completing an engineering loop from data construction, ChatML template conversion, 4-bit quantized loading, SFT finetuning, Adapter saving, to reload-based inference validation.

My Responsibilities

As the independent project owner, I designed and drove the vertical finetuning validation for an 8B open-source model around the QuanWen.Art support-agent scenario. I defined the training goal, scoped Chinese support and identity-QA data, planned the low-cost GPU experiment path, and completed the feasibility loop from training validation to future integration with the site's support corpus.

Technical Stack

Hugging Face ecosystem on dual T4 GPUs — SFT / QLoRA fine-tuning validation for an 8B model.

Data layer builds Chinese instruction/input/output samples covering identity Q&A, technical explanation, product copywriting, and customer-service replies.
Transformers + Datasets + PEFT + TRL + bitsandbytes form the full fine-tuning pipeline; ChatML template unifies the conversation format.
BitsAndBytesConfig uses 4bit NF4 + double quantization to reduce VRAM pressure; LoRA is injected into q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj modules.
paged_adamw_8bit optimizer, gradient accumulation, and k-bit training adapt to low-cost GPU experimentation environments.
Adapter save, compress, reload, and inference validation form a complete closed loop.

Key Milestones

Completed the Train -> Save -> Package -> Reload -> Inference finetuning loop for Qwen3-8B in a low-cost cloud GPU environment, producing a reusable 8B-level SFT / QLoRA engineering template. The project validates the feasibility of Chinese vertical-scenario finetuning under limited VRAM and establishes the engineering foundation for later QuanWen.Art support-corpus integration, private Q&A expansion, and a site-specific support agent.

Qwen3-8B SFT QLoRA PEFT Hugging Face BitsAndBytes ChatML Support Agent

Kaggle

GEO / SEO Readability Infrastructure - AI Search and Structured Indexing

GEO SEO llms.txt Structured Data

Project Background

A personal website should not only be a visual page for human visitors; it should also become a public content node that search engines, AI search systems, and model-training crawlers can understand reliably. This project upgrades QuanWen.Art from page-level presentation into a machine-readable site with robots.txt, sitemap.xml, llms.txt, and JSON-LD structured data.

My Responsibilities

Owned the overall governance of QuanWen.Art's external discoverability and machine readability, defining the boundaries between public content, private paths, search-engine crawling, and AI crawler access. I led the loop from policy design and semantic content organization to production verification, upgrading the site from a visitor-facing portfolio into a public knowledge node that search engines, answer engines, and LLM systems can understand consistently.

Technical Stack

FastAPI dynamically generates robots.txt, sitemap.xml, llms.txt, and JSON-LD, aligning public knowledge nodes with the search and AI ecosystem.

Route layer provides /robots.txt, /sitemap.xml, /llms.txt, and other public entrypoints; sitemap auto-outputs the homepage, photo list, progress pages, and published photo detail pages.
robots.txt explicitly allows Google, Bing, Baidu, Sogou, 360, Shenma search engines, plus AI crawlers: OpenAI, Perplexity, Applebot, Google-Extended, CCBot.
Simultaneously blocks /admin, /api, /passport, /partial, /uploads to prevent accidental indexing of private paths.
llms.txt describes site identity, theme scope, representative projects, and AI training/citation policy in Markdown.
JSON-LD uses Person, WebSite, ProfilePage, CollectionPage, CreativeWork, and ItemList from Schema.org to establish entity relationships.
pytest covers robots, sitemap, llms.txt, and structured-data output for homepage/photos/progress pages.

Key Milestones

This upgrade turns QuanWen.Art from a portfolio surface into a public knowledge node that machines can interpret. Search engines can follow sitemap-driven discovery; AI systems can use llms.txt to understand topic scope and authorization boundaries; structured data helps search and answer engines connect the author, website, projects, and content collections. Verification covers robots, sitemap, llms.txt, and JSON-LD, with production endpoints returning HTTP 200 after service restart.

FastAPI robots.txt sitemap.xml llms.txt JSON-LD Schema.org AI Crawler pytest

View Dev Log llms.txt sitemap.xml

Personality Prediction and User Behavior Insights

Data Sci Scholarly

Project Background

Built a machine-learning analysis pipeline for user behavior insights, evaluating the value of personality prediction for user segmentation and operational strategy support.

My Responsibilities

Owned the translation of a research-style personality prediction question into a verifiable data-analysis task, defining the scope for data processing, feature interpretation, model comparison, and result communication. I focused on turning model outputs into business-readable user segmentation and behavioral insights for later CRM tagging, A/B testing, and personalized outreach discussion.

Technical Stack

Pandas + Scikit-learn + XGBoost build a personality-prediction analysis pipeline from data cleaning to classification and interpretation.

The data layer handles missing-value treatment, outlier filtering, feature encoding, and standardization.
XGBoost + Scikit-learn Pipeline organize training and validation; GridSearchCV performs hyperparameter search and k-fold cross-validation.
Standardization and one-hot encoding are placed inside the pipeline to reduce data-preprocessing leakage risk.
Model results are translated into discussion foundations for CRM tagging, A/B testing, and personalized outreach.

Key Milestones

Developed a complete analysis workflow from raw data to interpretable model outputs, providing a foundation for further discussion of CRM tagging, A/B testing, and personalized outreach strategies.

Pandas / Numpy XGBoost Scikit-learn Data Visualization

View Dev Log

Multi-Source Web Collection and Public Sentiment Engine

Data Eng Scalable Pipeline

Project Background

Built a collection engine for video comments and specialist news, continuously scraping, cleaning, and producing structured output for policy research and public sentiment monitoring.

My Responsibilities

Owned the conversion of scattered web, comment, and news-data needs into a reusable data-collection task, defining collection scope, fields, cleaning rules, and delivery format. The goal was to move one-off manual gathering into a sustainable data-engineering workflow for policy research, public-opinion observation, and later sentiment analysis.

Technical Stack

requests + asyncio + rvest build an async multi-source collection pipeline producing structured research materials.

asyncio concurrency pools improve multi-page collection throughput; random delays, User-Agent rotation, and Cookie session maintenance improve scraping stability.
The cleaning layer handles field extraction, deduplication, format standardization, and anomaly filtering; output fields and cleaning specifications are standardized upfront.
Final output is structured as CSV / table files ready for downstream sentiment analysis and policy research reuse.

Key Milestones

Created a complete workflow from scraping and cleaning to saving structured results, reducing manual organization of raw web content and leaving an extensible foundation for sentiment analysis and competitor monitoring.

Python / asyncio rvest Anti-Scraping Countermeasures Data Pipeline

View Dev Log

Yingdao RPA Automation

RPA Automation Productivity

Project Background

Introduced RPA automation to reduce manual effort in frequent, repetitive, error-prone workflows in Zhiying EPR.

My Responsibilities

Owned the automation redesign of repetitive back-office workflows in ERP operations, identifying high-frequency manual steps and reorganizing clicks, form maintenance, and result checks into automatable business nodes. I focused on process standardization, exception-scenario mapping, and operations-side usability validation so the RPA flow could serve daily work rather than remain a script demo.

Technical Stack

Yingdao RPA breaks repetitive backend operations into reusable automation nodes, with fallback handling for exceptions.

The flow layer decomposes login, menu navigation, form entry, and result confirmation into standard action nodes, avoiding one-off unmaintainable scripts.
The execution layer uses keyboard/mouse simulation and conditional triggers to complete backend operations; the exception fallback layer handles popups, timeouts, and retries.
Manual confirmation checkpoints are preserved at critical nodes to prevent automation errors from expanding risk.

Key Milestones

Turned repetitive back-office actions into reusable automation workflows, reducing execution variability from manual clicks and mistakes for continued use in stable business processes.

Yingdao RPA EPR Auto Click Workflow Orchestration Retry Handling

Amazon US FBM Deep-Listing

Strategy Ops Production Live

Project Background

For Amazon US FBM curated listings, the goal is to give new-product discovery, asset preparation, and listing cadence a firmer basis rather than relying on experience-led bulk listing.

My Responsibilities

Owned product-selection judgment, listing cadence, and asset-preparation coordination for Amazon US FBM deep-listing work, turning experience-driven listing into a more reviewable data-driven workflow. I connected market trends, keyword judgment, asset quality, and listing rhythm so operations strategy and the AI visual toolchain could cooperate more steadily.

Technical Stack

SellerSprite + Google Trends + AI visual toolchain form a closed loop for sourcing decisions, asset preparation, and listing-cadence review.

The data-judgment layer focuses on keyword heat, competition intensity, and margin potential; trend and keyword data constrain sourcing priorities.
The asset-production layer chains product-image optimization with AI scene-generation; the AI visual toolchain improves image-production consistency.
The operations-review layer tracks listing cadence, asset quality, and downstream performance, continuously refining sourcing rules and asset standards.

Key Milestones

Gradually developed more stable product-selection and launch rhythms, focusing on listing quality, consistent presentation, and efficient asset preparation for continuing long-term improvements.

SellerSprite Google Trends FBM Deep Listing Original White-Background Assets

Painter-Harness

The batch-generation and multi-turn editing foundation for image models

Core Project Iterative Dev Harness

Image-Model Harness Harness-level Multi-turn Editing Prefix-Cache Friendly

Project Background

Image models are stateless: one call produces one image, with no memory of batch progress, no knowledge of why the previous image failed, and no guarantee of consistency with references. Product-detail image production, however, is a long-running, multi-step, cross-session task.

Painter-Harness supplies the layer outside the model. Composition contracts, constraint scaffolding, model routing, externalized state, and failure recovery are handled by the runtime; the model only paints.

It natively supports AI image generation and multi-turn conversational editing through the Harness layer, without relying on the model to preserve a session. Any generated image can become a new baseline, accept additional references, and produce variants turn by turn while product anchors and copy constraints remain stable throughout the session.

Under the common definition Agent = Model + Harness, this project builds the second half for image models.

My Responsibilities

Designed and implemented this runtime layer. The focus is not prompt tuning, but contract definition, state placement, failure recovery, and permission boundaries, with responsibility for output consistency, protocol compatibility, quota scheduling, and replayable history.

Technical Stack

A four-layer structure with model calls enclosed by deterministic code.

Scaffold

Product anchoring, visible-copy allowlists, metadata no-render rules, and color fidelity are encoded as structured constraints and assembled on every call without relying on accumulated context. Composition is contracted as nodes, with each node becoming an independent task.

Runtime

Scans directories, creates tasks, assembles requests, dispatches models, collects outputs, and archives them to disk. Model-family differences are absorbed by one adapter layer: generation and editing share the same task abstraction, while unsupported optional parameters degrade step by step without interrupting the batch. Concurrency, spacing, timeouts, and quotas are scheduled centrally by the runtime.

Harness Capabilities

Multi-turn editing does not rely on native model session retention; the runtime owns it. Every image carries a session, while session state, references, and variant lineage persist outside the model. Each turn is rebuilt as an independent request from the baseline image, added references, and the current instruction. Models with single-turn APIs therefore gain multi-turn editing; resulting variants are stored, can become the current image, and remain resumable across sessions.

Assembly order is fixed as stable constraint prefix → variable node and instruction. Models supporting prefix caching can hit the cache directly, amortizing repeated constraint input and materially reducing generation cost across long batches and multi-turn edits.

State

Progress lives outside the runtime. Batch checkpoints skip completed nodes across runs; turn archives let additions, edits, and variant comparisons continue without rerunning the whole batch.

Guardrail

State and logs stream in real time. Failures retain error class, quota source, and retry history, and tasks can be cancelled at any time. High-cost generation is excluded from the automation tool catalog. Upscaling and size convergence are available as optional post-processing stages.

Key Milestones

The model stays the same; delivery determinism comes from the runtime. Directory contracts make inputs predictable, scaffolding keeps outputs consistent, checkpoints contain failures to one node, and session state preserves context during refinement.

Once multi-turn editing becomes a Harness capability, model choice is no longer constrained by native session support: single-turn models can enter the editing chain. Stable-prefix assembly lets cache-aware models further amortize repeated constraints. One structured node set can drive multilingual variants, while batch generation and multi-turn editing of a single image run through the same tooling layer.

Python 3.12 · FastAPI · Pydantic · Native JS Control Plane

Tool Scheduling · Concurrency Orchestration · Quota Pool · Retry and Degradation Multi-family Routing · Generation / Editing Channels · Parameter Compatibility Constraint Scaffold · Structured Node Contracts · Output Parsing and Validation Harness-level Multi-turn Editing · Variant Lineage · Reference Sessions Stable Prefix Assembly · Prefix Cache Hits · Single-turn Model Compatibility · Cost Amortization Externalized State · Batch Checkpoints · Memory-like State Observable Log Stream · Cancellable Tasks · Permission Convergence

AI Batch Design for International Trade: JSON Workflow Generator

Generate standard, highly controllable SD / Midjourney automated image-rendering JSON metadata in one click

Artistic Direction Style

Core Export Goods

My Responsibilities

Owned the productization of an AI image-prompt workflow demo, turning style, product, composition, and rendering parameters for foreign-trade product images into configurable Prompt JSON. The focus was moving image generation from hand-written prompts into a structured, reusable, and demonstrable interface for understanding parameterized visual production.

Technical Stack

Abstracts foreign-trade product-detail image generation into a structured Prompt JSON, making style, product attributes, composition, and rendering parameters configurable, previewable, and reusable.

The parameter layer manages product type, visual style, selling points, background, camera/lens, and output specs.
Form parameters are assembled into structured Prompt JSON; the output preserves field constraints for later integration with ComfyUI, Flux, or other image-generation workflows.
Instant preview strengthens the understanding of "parameter change → visual intent change," reducing the cost of reusing and modifying hand-written prompts.

Blueprint copied to clipboard! Ready for Stable Diffusion / Midjourney automation.

Generating JSON system blueprint...

About Me — Narrative Poetry Meets Science

Core Competencies

01. Trade Research & Platform Ops

02. Full-Stack Eng & Web Architecture

03. Agent Eng & DevOps Hub

04. Open Source & Front-End Aesthetics

Atelier Showcase & Laboratories

General Tool Platform (AiNative-Tools)

Project Background

My Responsibilities

Technical Stack

Key Milestones

Lite Studio Client

Project Background

My Responsibilities

Technical Stack

Key Milestones

OpenClaw Sentinel

Project Background

My Responsibilities

Technical Stack

Key Milestones

MiMo Vision MCP

Project Background

My Responsibilities

Technical Stack

Key Milestones

QCode Toolkit (SVG QR Code)

Project Background

My Responsibilities

Technical Stack

Key Milestones

QuanWen.art

Project Background

My Responsibilities

Technical Stack

Key Milestones

Qwen3-8B Chinese Support Agent Finetuning

Project Background

My Responsibilities

Technical Stack

Key Milestones

GEO / SEO Readability Infrastructure - AI Search and Structured Indexing

Project Background

My Responsibilities

Technical Stack

Key Milestones

Personality Prediction and User Behavior Insights

Project Background

My Responsibilities

Technical Stack

Key Milestones

Multi-Source Web Collection and Public Sentiment Engine

Project Background

My Responsibilities

Technical Stack

Key Milestones

Yingdao RPA Automation

Project Background

My Responsibilities

Technical Stack

Key Milestones

Amazon US FBM Deep-Listing

Project Background

My Responsibilities

Technical Stack

Key Milestones

Painter-Harness

Project Background

My Responsibilities

Technical Stack

Scaffold

Runtime

Harness Capabilities

State

Guardrail

Key Milestones

AI Batch Design for International Trade: JSON Workflow Generator

My Responsibilities