Tea-market Deep Research and interactive visualization workbench
Sep 2026
A Deep Research and interactive-visualization workbench for tea-market research. It defines the market scope, entity hierarchy, scale/price/growth metrics, and evidence contract, then turns the research workflow into structured data with sources, status, and missing-state notes before mapping it into an explorable ECharts market view. It supports vendor and consumer-label perspectives, brand tiers, representative price, retail sales and growth, product-line detail, Best SKU/new records, and reproducible URL filters. The current release is a clearly labeled Mock used to validate the research–data–visualization loop.
View demo ↗GEO experimentation and brand intelligence workbench
Concept: Dec 2025 · Development: Apr 2026–present
A GEO platform conceived in December 2025, started in April 2026, and developed continuously since then. After exploring an RPA approach, I selected a BrowserUse-driven browser-sampling path and built the workflow from GEO experiment setup and cross-platform sampling to result persistence, brand perception and citation analysis, evidence chains, customer-ready reports, and ongoing monitoring. It is a repeatable experimentation and explainable-evaluation workbench, live at geo.vationx.ai.
View demo ↗Medical review Agent and spreadsheet workbench
Mar–Aug 2026
An AE medical-review assistant for clinical-trial data. It evolved from an Excel add-in through a LobeHub + Univer + Skill workspace into a Pi Agent + Skill and local self-harness evaluation loop. Subject-level evidence packages, controlled retrieval, record-level output, and development/holdout isolation helped reach 70%+ record-level recall on a controlled validation set without large-scale row-by-row labeling.
View case study ↗AI product website Hero interaction prototype
Jun 2026
A responsive Hero prototype that turns an AI media-recommendation workflow into a product narrative. CSS- and GSAP-driven versions were used to explore the balance between brand expression, task visualization, and motion.
View demo ↗AI media recommendation and data-governance workbench
Jan–Jun 2026
A product project delivered from January to June 2026. Its product demo turns brief interpretation, candidate filtering, explainable recommendations, outreach, and export approval into an operational workflow, alongside data intake, review, publishing, feedback remediation, and sensitive-action approval.
View demo ↗Enterprise AI platform assessment and architecture blueprint
Apr–May 2026
A client-facing enterprise AI platform selection study comparing Tencent ADP, Alibaba Bailian, and Volcano Coze across agent capabilities, MCP and systems integration, security governance, cost, and delivery. The work produced a scenario-capability map, data-security comparison, POC recommendations, and a reusable three-layer architecture blueprint. The public page presents the report outline, assessment method, and architecture summary.
View report outline ↗Multimodal content generation studio
Jan–Mar 2026
A multimodal content product project delivered from January to March 2026. It brings PDF, PPTX, DOCX, and XLSX source materials into one workspace, extracts traceable knowledge structures, and generates training assessments with source-PDF review, preview, answering, scoring, and regeneration flows. The linked page is a product-demo recreation from June.
View demo ↗E-commerce and finance integration
Jan 2023–Feb 2025
Connected products, orders, settlement, supplier reconciliation, and financial compliance data. Customer, contract, organization, and product master data reached full online coverage.
Employee AI assistant
Jan 2022–Dec 2024
An enterprise HR, administration, and IT knowledge assistant using RAG and large-model workflows. It served 2,000+ employees through WeCom, raising FAQ accuracy by about 40% and reaching about 80% daily coverage.
HR digital recognition platform
Jan 2022–Dec 2024
An employee-experience SaaS product serving multiple large customers, later extended with large-model self-service Q&A.
Social security policy Copilot
Sep–Oct 2023
A product validation from hackathon to PoC. Prompt design, Copilot interaction, and 500+ real conversation reviews reduced the average response workload for specialists by about 60%.