AI decision systems / Agent workflow / Field data products
付皓
AI decision systems, Agent workflow, and field-tested data products
I turn ambiguous field problems into data definitions, model workflows, system interfaces, and delivery artifacts. My strongest evidence now spans DJI x National Energy Group industrial drone inspection, maritime battery thermal-runaway monitoring, medical AI evaluation and modeling, a multi-agent study workbench, Tsinghua satellite-ground scheduling, and Guanlan ERP / WeChat Mini Program delivery.
Claims matter when the evidence remains inspectable.
Operating notes
Recruiter-safe proof for industrial AI, AIoT, medical evaluation, Agent workflow, scheduling, and enterprise delivery. Private logs, customer data, and application strategy stay out of the public site.
From physical constraints to decision loops
Drone inspection, battery safety, AIoT edge-cloud design, and risk-state machines
Core projects
Casefiles arranged by real operating problem, not by tool list.
Each case starts with ambiguity, then resolves into data definitions, architecture, workflow, and public-safe evidence.
Real constraints. Clear boundaries. Evidence before claims.
Flight queue
Agent Workflow
Supports Agent product, evaluation workflow, tool-use orchestration, and AI learning-product roles.
Challenge
Long lecture decks and mixed-format course material lose context when a model reads everything in one pass. A useful study agent needs planning, retrieval, page-level explanation, quality gates, and recoverable failure handling.
Architecture
Built a LangGraph Planner / Reasoner workflow with document preprocessing, layout analysis, local retrieval, page-by-page explanation, and Tool-Use interfaces. The system keeps global memory for long-range planning, then executes chapter and page tasks with code-controlled quality scoring before output.
Ownership
Stack
Outcome
A working study-agent workbench that turns fragile prompt-only reading into an inspectable workflow with self-correction and process-supervision data collection.
Define the boundary
Build the loop
Keep it habitable
Public proof
Sanitized case pages that show what each project proves, where the boundaries are, and which role signals the evidence supports.
No raw logs, private dashboards, credentials, or customer data.
Agent Workflow
A study-agent workbench that decomposes long course material into planned, tool-assisted, page-level learning workflows.
Business Delivery
A real business digitization case spanning ERP-style data management, WeChat Mini Program, website, and e-commerce order integration.
Scheduling System
A ground scheduling system for a 400+ person optical satellite photo event with real spacing, feasibility, and task-card constraints.
Medical AI Eval
A cross-institution medical AI project covering NIH R01 text evaluation and osteoporosis modeling with migration validation.
AIoT Risk System
An AIoT edge-cloud monitoring project for EV battery thermal-runaway risk during maritime transport.
Industrial AI
An early industrial drone-inspection project for high-altitude confined-space equipment checks in thermal power plants.
Field structure
The value is the combination: engineering, Agent workflow, modeling, AIoT, and field delivery.
Capabilities are mapped to evidence, not left as tool buzzwords.
Capability node
I can build the data, API, deployment, and interface pieces needed to make an idea testable.
Capability node
My Agent work focuses on task decomposition, tool use, quality gates, and observable execution rather than prompt theater.
Capability node
I translate messy research or clinical data into feature definitions, model comparisons, and reviewable decisions.
Capability node
The recurring pattern is physical or business ambiguity becoming a usable workflow with clear boundaries.
Measured history
Education, awards, and project milestones that explain the current operating profile.
Dates are aligned to the latest July 2026 resume.
Education
M.S. in Applied Data Science and Statistics, with coursework and projects across statistical modeling, machine learning, text evaluation, and medical data modeling.
Mathematics and Computer Science undergraduate background; Dean's List recognition in 2022.
Awards
Advanced on site in Rotterdam to the October 2026 Innovate Together Competition finals; global selection ratio 10 / 44.
First-place final result and best practice achievement report for the maritime battery monitoring project.
Team captain; second prize in the CSG venue among 12 finalist teams and UNITAR SDG Open Hack X-lab second prize.
Awarded by China Electric Power Technology Market Association for the DJI x National Energy Group industrial drone inspection project.
Recognized for QC group activity results related to the industrial inspection project.
Milestones
LangGraph Planner / Reasoner workflow with Tool-Use, retrieval, LLM/VLM routing, quality gates, and process-supervision traces.
Delivered ERP-style data management, internal Mini Program, public website, and e-commerce order integration.
Built a 400+ participant ground scheduling and matrix layout system for an optical satellite photo event.
Worked on NIH R01 text evaluation and osteoporosis cross-institution modeling with UTHSC and St. Jude collaborators.
AIoT edge-cloud risk monitoring project with patent authorization, ITC advancement, Tsinghua gold, and UNITAR second-prize results.
High-altitude confined-space inspection method with visual thresholds, 100°C early-entry workflow, innovation gold award, QC third prize, and patent disclosure.