_{Click the preview to play the full demo (muted) · 10.75 s · 1920x1080 · 60 fps · 400-690 nm.}

This is an earlier private hyperspectral reconstruction prototype. The public loop and full demo show a desktop workflow for camera acquisition, lightweight deep-learning reconstruction, 30-band spectral display across 400-690 nm, per-point spectral curves, and result saving.

I was responsible for the lightweight deep-learning reconstruction and dataset work, plus the camera acquisition, desktop UI, spectral visualization, and result-saving modules. Code and data remain private because of research and industry-collaboration constraints.

I build reliable AI applications end to end, connecting product interfaces, agent workflows, APIs, and data systems. Computational imaging is my technical edge.

AI applications   frontend + client systems   agent workflows   computational imaging

I turn AI capabilities into products people can use: streaming interfaces, local client workflows, orchestration, APIs, and data delivery. My main lane is AI full-stack engineering, with computational imaging and open-source collaboration adding technical depth.

• AI application engineering: turning LLM and agent capabilities into product flows across UI, orchestration, APIs, and data.
• Frontend and client engineering: building streaming interfaces, desktop workflows, device-facing controls, and cross-platform behavior.
• Agent workflows and data systems: working with RAG, DAG execution, model routing, MCP, SSE, RBAC, state, and storage.
• Computational imaging: turning reconstruction research into usable software through camera acquisition, spectral visualization, and result handling.

• Interface: HTML5, CSS3, JavaScript, TypeScript, React, React Flow, Zustand, Vite, Webpack, and Rspack.
• AI applications: LLM agents, RAG, LangChain, LangGraph, MCP tool calling, model routing, SSE, and WebSocket workflows.
• Services and data: Node.js, NestJS, Python, PyTorch, PostgreSQL, pgvector, Redis, DAG execution, RBAC, snapshots, and structured diff or rollback.
• Client delivery: Rust, Tauri, Git, Windows tooling, PyInstaller, camera workflows, and desktop result handling.

My coding-agent toolkit includes Kimi Code, Codex, Pi, Claude Code, Hermes, OpenCode, and ZCode. I also follow DeepSeek Harness within the broader agent tooling ecosystem. I use these tools inside a human-led loop: understand the codebase, implement the change, test edge cases, document decisions, and review the result.

I work on real issues across AI agent systems, frontend ecosystems, desktop and web tooling, and civic data. As of 2026-09-17, 94 PRs are merged across 32 upstream projects, with 78 more in review.

Recently merged across major Chinese AI agent and frontend ecosystems:

• 🐜 ant-design/ant-design#59173 · ⭐ 99.5k — fixed the Chinese contribution guide to link the English locale list in its i18n step, so contributors update both locale lists.
• 🔍 alibaba/open-code-review · ⭐ 33.5k — added Pug template support to the review allowlist (#1114), then made report writes atomic so interrupted runs cannot leave truncated output (#1160).
• 🤖 agentscope-ai/agentscope#2434 · ⭐ 31.9k — Uvicorn hot reload forced a SelectorEventLoop on Windows and broke subprocess execution; disabling reload on Windows keeps a compatible event loop.
• 💬 QwenLM/qwen-code#11300 · ⭐ 27.9k — kept branch commits made before a failing post-checkout hook, so a failed hook no longer discards committed work.
• ⚛️ alibaba/hooks#2956 · ⭐ 15.0k — kept useLatest and useUnmountedRef returns as non-nullable MutableRefObject across the supported React type versions, backed by declaration-emission regression tests.
• 🎨 DouyinFE/semi-design#3352 · ⭐ 10.4k — made ESC close only the top-most modal, so stacked modals no longer collapse together.
• 🍒 Tencent/cherry-markdown#1871 · ⭐ 4.9k — guarded pending image load and error callbacks after previewer destroy, so late events no longer touch torn-down state.
• 🧩 Tencent/tdesign-vue-next · ⭐ 2.2k — restored the Dropdown onClick option type (#6959) and made Menu handle functional submenu parents (#6967).

Merged in international frontend and tooling ecosystems:

• ⚡ chakra-ui/chakra-ui#10979 · ⭐ 40.6k — gave the file-upload delete trigger a proper disabled style.
• 📦 pnpm/pnpm#14708 · ⭐ 36.5k — stopped emitting empty package authors in SBOM documents.

28 merged PRs in grafana/grafana-pathfinder-app · ⭐ 19 — sustained features, fixes, and hardening that span Grafana's interactive learning app end to end: block-editor bulk actions, coda VM lifecycle, docs retrieval, and lint plus e2e coverage.

27 merged PRs in open-city-ai/haidian — selected deep dives:

_{…plus single PRs across parcel, desktop/desktop, payloadcms, continue, eslint, NextChat, OpenViking, and other ecosystems — the auto-refreshed 📊 Full footprint below tracks every upstream project.}

• Master's student — Electronic Engineering and Information Photonics, BUPT
• Undergraduate background — Network Engineering, BUPT

dvd@linux.do · z1403594118@gmail.com