Zedong Jiang

1688, Taotian Group

• Sole development owner for AI applications, from requirements and solution design to launch validation and iteration.
• Built an end-to-end AI interview system: job posting, resume parsing and matching, screening, adaptive interviews and evaluation reports, with AI and human checkers.
• Converted business requirements into agent role contracts, evaluation blueprints and executable tests. Assessed sandboxed Candidate Agents using checkpoints, deterministic checkers, LLM reviewers and execution traces.
• Built a rubric-based LLM-as-a-Judge checker for Zeus, an executive decision-support agent. Evaluated precision, recall, F1 and missed P0 issues through case-based regression testing.

Agent evaluation · Tool calling · LLM-as-a-Judge

3Chat.ai

Built a Python workflow connecting product-data collection, background removal, AI background generation, text layout and rendering for e-commerce images. For standardized SKUs, reduced processing time from about 10 minutes to 1 minute per image.

Python · AI image workflows · Automation

SecondMe

• Contributed to the A2A agent pipeline: profiles, long-term memory, prompt rendering, model calls, tool execution and result persistence.
• Integrated external APIs and MCP services as agent-callable skills, with argument validation, routing, permission allowlists, dynamic publishing and rollback.
• Replayed requests across Claude and GPT; used prompt A/B comparisons, Langfuse, CLS and staging E2E tests to diagnose context and state-persistence issues.

A2A · MCP · Skills · E2E

Xinhua Zhiyun

Developed geographic and entity-visualization components for the Gewu B2B platform: POI search, advanced tag filtering, route display and editing, layers and interactive pop-ups. Independently delivered multiple releases.

TypeScript · G6 · B2B

Vela

Customer-service, sales and operations agents for cross-border e-commerce. Built smoke tests from simulated inquiry and knowledge retrieval through human confirmation and result persistence; built a locally runnable agent-worker platform with configurable skills, tool routing and permission allowlists.

Agents · Knowledge bases · Smoke testing

TripPup

Independently delivered an iOS app and Next.js backend for itinerary planning, menu understanding, nearby discovery and document workflows. Built context injection, OCR, function calling, model fallback, rate limiting and error recovery.

SwiftUI · Next.js · Supabase

PRD Genius

A staged workflow for requirements, user stories, competitive analysis, flowcharts, mind maps and runnable HTML prototypes. Editable JSON Schema outputs, refinement, undo/redo and HTML/PDF export.

React · TypeScript · Gemini 2.5 Flash

AI Nexus

A bilingual directory and community workflow with LLM recommendations and metadata autofill. Built uploaded-file builds, Blob distribution and moderated subdomain publishing.

React · Supabase · Vercel Serverless

Talent Excavator

A Socratic conversation interface inspired by Gallup, flow theory and Jungian psychology, with streaming dialogue and printable personal talent reports.

Next.js · Gemini Streaming

AI & Agent Engineering

Agent workflows · Function calling · MCP · RAG · JSON Schema · OCR · Context and memory · Model fallback

Evaluation & Debugging

Rubric-based evaluation · LLM-as-a-Judge · Precision / Recall / F1 · Langfuse · CLS · Prompt A/B · E2E · Smoke tests

Full-stack Development

TypeScript · JavaScript · Swift · Python · SQL · React · Next.js · SwiftUI · Supabase · REST APIs

Delivery & Reliability

Vercel Serverless · Docker · esbuild · Git · Rate limiting · Error recovery · App Store delivery