Led stack selection and 0-to-1 architecture for AI Media Studio, taking a multimodal AI creation platform from blank slate to production release.

Established team-wide AI Coding practice around spec-driven development and agent collaboration, raising delivery speed and review quality across the team.

Built the streaming chat foundation on the UI Message Stream protocol with server-side snapshots, so sessions survive a dropped connection and resume without losing context.

Delivered the core capability set of open-source desktop app 302 AI Studio — a coding-agent runtime that runs in local containers or managed cloud sandboxes, with full MCP support for plugging in external tool servers.

Automated the app's internal service layer with a build-time code generator, removing hand-written IPC boilerplate across 30+ main-process services and making main/renderer contracts type-checked end to end.

Kept the app stable and responsive across long multi-session workloads — memory governance suspends inactive background renderers to prevent out-of-memory crashes, and stream handling was rebuilt around viewport visibility.

Improved one-click recolor — the suite's top-three feature — with a reworked color-replacement pipeline that cut average processing time from 330ms to 60ms.

Built the interactive inpainting canvas, bringing lasso selection, undo/redo, and hover-based segment highlighting into the editor for manual retouching on top of AI generation.

Orchestrated Stable Diffusion workflows and prompt tuning with TensorRT and XFormers acceleration, lifting image generation throughput by roughly 60%–80%.

Contributed to 0-to-1 delivery of an LLM-driven RPA desktop client that compiles natural-language instructions into runnable automation code.

Built the automation runtime on Electron and Puppeteer, reaching native OS capabilities through cross-language DLL calls so generated scripts could drive desktop software.

Implemented headless session verification and co-developed the Wizard DSL-to-JavaScript transpiler, backed by unit tests over the generated code paths.

Published a widely cited review on 'Optimization strategies of fruit detection to overcome the challenge of unstructured background in field orchard environment' in Precision Agriculture (IF 7.6, CAS Tier-1 Top), cited 160+ times.

Published 'An Enhanced Cycle Generative Adversarial Network Approach for Nighttime Pineapple Detection of Automated Harvesting Robots' in Agronomy (IF 4.1, JCR Q1), cited 30+ times — advancing nighttime robotic harvesting vision.