Projects

A Python multi-agent decision system that connects intent understanding, clarification, candidate generation, risk review, and explanation agents with a deterministic ranking engine for stable, testable recommendations.

• ▸ A multi-agent workflow for intent, understanding, clarification, candidates, planning, review, explanation, risk, and evaluation
• ▸ A deterministic ranking engine with hard-constraint filtering and seven-dimensional scoring
• ▸ FastAPI plus a local web UI, with rule agents that work without an API key

Built an evidence-first Agent evaluation workbench with 60 samples across 10 task types, a five-dimensional rubric, A/B comparison, bad-case analysis, and quality, cost, P95 latency, and safety gates.

• ▸ A 60-sample Golden Set spanning 10 task types instead of a few curated demos
• ▸ A five-dimensional rubric, A/B comparison, and bad-case analysis for explainable choices
• ▸ Quality, cost, P95 latency, and safety gates with JSON artifacts and an HTML report

A health-information product with quick checks, guided practice, and progress review, helping users understand evidence, practice judgment, and build their own reasoning habits.

• ▸ Designed three product modes: Quick Check, Guided Practice, and Progress Review
• ▸ Built a personalized loop around scenario judgment, instant feedback, and transfer practice
• ▸ Used RAG for traceable evidence and defined a three-group product validation plan

Used structured specifications to organize requirements and interaction design, then worked with coding agents across frontend and backend implementation, integration testing, and launch coordination.

• ▸ Covered requirement scoping, interaction design, and structured specifications
• ▸ Contributed to JavaScript frontend and Go backend implementation
• ▸ Worked with engineering and QA through integration, fixes, and launch

Turned personal positioning, project evidence, and content-maintenance needs into an evolving bilingual site, using Research → Plan → Implement, build checks, and deployment as an AI Coding loop.

• ▸ Designed a bilingual portfolio architecture around positioning, project evidence, and reading paths
• ▸ Used Research → Plan → Implement to constrain AI Coding across content, interaction, templates, and deployment
• ▸ Validated delivery through Hugo builds, release checks, Vercel, and GitHub Pages workflows

Independently researched a manufacturing equipment-management problem, modeled its workflows and UML, designed the platform and database, and adapted an open-source codebase into a Vue + Spring Boot system.

• ▸ Identified fragmented equipment data and imprecise maintenance planning through an enterprise internship and research
• ▸ Independently completed business flows, UML models, module architecture, database design, and system architecture
• ▸ Adapted an open-source codebase into a Vue + Spring Boot system without AI-assisted coding

Owned competitive analysis, business-model design, and financial forecasting for an air-ground delivery concept, connecting differentiation to revenue paths and funding assumptions.

• ▸ Compared competitors across precision, operating range, automation, digital capabilities, and price
• ▸ Designed a business model linking R&D, production, platform services, and downstream applications
• ▸ Built five-year revenue, profit, expense, financing, and fund-allocation forecasts

Studied online music payment through user behavior and consumer psychology, validated the survey instrument, and used a random forest to turn payment drivers into platform recommendations.

• ▸ Proposed the selected user-centered topic and led questionnaire design, pilot testing, and the formal survey
• ▸ Ranked seven payment drivers with a random forest; catalog and sound quality contributed about 59.69% of total importance
• ▸ Wrote the related conclusions and recommendations; the project received a national third prize

A lightweight food-and-drink picker that narrows options by occasion and flavor, supports temporary exclusions, and uses rating-weighted randomness to turn repeated comparison into an acceptable decision.

• ▸ Turned my recurring what-to-eat and what-to-drink indecision into a minimal decision loop
• ▸ Narrows candidates by occasion, flavor, and temporary exclusions before a rating-weighted draw
• ▸ Supports custom pools, categories, and the latest 20 results, with preferences stored locally

A lightweight reading concept that matches short books to the time, mood, and context of a commute, lunch break, or evening wind-down.

• ▸ Uses available reading time and current state as recommendation inputs
• ▸ Defines complementary ReadNow and CanonLit scores
• ▸ Reduces search effort through curated lists and standardized book cards

A Provincial First Prize project for Problem C of the 2022 CUMCM, analyzing weathering, composition changes, and type identification for ancient glass artifacts through statistical modeling and classification.

• ▸ Led code implementation for random forest, Bayes, multivariate regression, and sensitivity analysis while supporting model development
• ▸ The team combined statistical tests, regression, clustering, and classification to analyze ancient glass and unknown samples
• ▸ Won a Provincial First Prize in Problem C of the 2022 CUMCM

Built layered models for forest carbon sequestration, harvesting economics, and water conservation, then applied a multi-objective management plan to larch and Mongolian oak forests in eastern Liaoning.

• ▸ Primarily responsible for project modeling, with a limited contribution to paper writing; the team received an Honorable Mention
• ▸ Separated carbon sequestration by living trees and wood products and used a Richards growth curve for long-term estimation
• ▸ Applied AHP to five water-conservation factors while accounting for harvesting revenue and long-term management costs