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

What the demo shows

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

About

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.

Engineering focus

• 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.

Tech stack

• 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.

AI coding workflow

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.

 Kimi Code   Codex   Pi   Claude Code

 Hermes   OpenCode   ZCode   DeepSeek Harness

🌐 Open-source contributions

I work on real issues across AI agent systems, frontend ecosystems, desktop and web tooling, and civic data. As of 2026-09-01, 32 PRs are merged across six upstream projects, with 24 more in review.

🏆 Merged proof

Recently merged across major AI agent and frontend ecosystems:

• 🐜 ant-design/ant-design#59173 · ⭐ 99.3k — fixed the Chinese contribution guide to link the English locale list in its i18n step, so contributors update both locale lists.
• 🤖 agentscope-ai/agentscope#2434 · ⭐ 30.3k — Uvicorn hot reload forced a SelectorEventLoop on Windows and broke subprocess execution; disabling reload on Windows keeps a compatible event loop.
• 🔍 alibaba/open-code-review#1114 · ⭐ 21.8k — added Pug template support to the review allowlist, with a dedicated rule covering escaping contexts, attribute spreads, and URL sinks.
• ⚛️ 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.
• 🍒 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.

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

🥇 Complete delivery · bilingual formal submission

Problem: A civic-data submission needed a complete, reviewable package rather than an unsupported concept.

Contribution: I delivered a scoped 41-file bilingual package spanning narrative, offline HTML/PDF, GeoJSON, metrics, sources, a manifest, and deterministic self-checks, while marking provisional geometry and non-endorsement boundaries.

PR: open-city-ai/haidian#2540 · merged for repository intake, not gallery publication or government endorsement.

🔎 Site provenance

Problem: Provisional site inputs lacked consistent dates, hashes, licence and use boundaries, while an empty constraints layer could be mistaken for proof that no controls existed.

Contribution: I added a built-ins-only audit for four submitted geometries, recorded six missing official-control layers, and made replacement triggers and permitted uses traceable across the package. The merged change passed 18/18 audit checks.

PR: open-city-ai/haidian#3525

🪟 Windows portability · import and worker locking

Problem: An unconditional fcntl import stopped two test modules from being collected on Windows, and worker-lock contention needed consistent cross-platform semantics.

Contribution: I reproduced the failure, introduced the guarded POSIX and Windows lock path, and added the initial regression. Maintainer review then corrected the production file-open path and strengthened its test; the previously uncollectable target suite passed 21/21 tests.

PR: open-city-ai/haidian#3859

📄 Deterministic artifacts · LF across platforms

Problem: Three writer paths emitted CRLF on Windows, creating whole-file diff noise and breaking byte-hash consistency after line-ending normalization.

Contribution: I made the writers emit LF bytes and added direct plus end-to-end regressions. After review exposed a Python 3.9 compatibility issue, I revised the implementation and revalidated both flows.

PR: open-city-ai/haidian#3915

⚖️ Review semantics · blockers versus future follow-ups

Problem: A flat review-action list converted future conditions into immediate change requests, creating an intake loop that compliant edits could not close.

Contribution: I introduced explicit non-blocking follow-ups with trigger and owner fields, invalidated legacy review caches, and kept current or malformed repairs fail-closed with focused regression coverage.

PR: open-city-ai/haidian#3982; the original blocked head was re-reviewed and merged without weakening existing gates.

🗺️ Contribution map

Area	Project	Status	Selected work
🏙️ Civic data delivery	open-city-ai/haidian	28 PRs · 27 merged	Structured submissions, geometry sources, deterministic artifacts, review semantics, and a CJK font screenshot regression
⚛️ Frontend ecosystems	ant-design/ant-design · alibaba/hooks · Tencent/cherry-markdown	3 merged	Contribution-guide link repair, non-nullable ref types, and previewer lifecycle cleanup
🔍 Code review tooling	alibaba/open-code-review	1 merged · 1 open	Pug template support merged; Jinja template support in review
🤖 AI agent systems	agentscope-ai/agentscope	2 PRs · 1 merged	Windows subprocess behavior in an AI agent example
📱 Cross-end frontend	NervJS/taro	6 open	Component rendering, platform API typings, and Vite runner CSS behavior
🖥️ Agent desktop UI	bytedance/UI-TARS-desktop	5 open	Agent server streaming, MCP initialization, device actions, and workspace search ordering
🧠 Agent memory & Node.js	TencentCloud/TencentDB-Agent-Memory	3 open	Proxy preservation, SQLite filtering, and self-healing store state
💻 Desktop UX	desktop/desktop	1 open	React and TypeScript pull-request state and suggestion behavior
📦 Web tooling	parcel-bundler/parcel	1 open	Correct UTF-8 BOM JSON parsing in the resolver path
🧭 Search & coordinates	organicmaps/organicmaps	1 open	C++ search behavior for space-separated DMS coordinates

🚀 In review now

🤖 AI agent systems · 9 open PRs

• TencentCloud/TencentDB-Agent-Memory#1171: keep Node's global proxy dispatcher intact while loading undici@8.
• TencentCloud/TencentDB-Agent-Memory#1172: self-heal the store-init cache after failed initialization or a closed store.
• TencentCloud/TencentDB-Agent-Memory#1173: honor recordIds filtering in the SQLite L1 query path.
• bytedance/UI-TARS-desktop#1957: release the exclusive slot when agent-server stream startup fails.
• bytedance/UI-TARS-desktop#1958: await asynchronous ADB device actions.
• bytedance/UI-TARS-desktop#1959: add default search exclusions to the MCP filesystem server.
• bytedance/UI-TARS-desktop#1960: surface MCP agent initialization failures.
• bytedance/UI-TARS-desktop#1961: prioritize source workspace results in the contextual selector.
• QwenLM/qwen-code#10580: localize untitled session search in the VS Code integration.

📱 Cross-end frontend · 6 open PRs

• NervJS/taro#19491: center aspectFit images horizontally.
• NervJS/taro#19492: keep global page styles when syncing CSS in the Vite runner.
• NervJS/taro#19493: add the WeChat phone-number quota toast prop to Button and its template.
• NervJS/taro#19494: mark holdKeyboard for Baidu and pass hold-keyboard through the swan templates.
• NervJS/taro#19495: fix self-contradictory generic constraints and missing RequestParams in the request typings.
• NervJS/taro#19496: add the missing requestCommonPayment typings for WeChat payment.

🛠️ Frontend, client, and tooling · 8 open PRs

• alibaba/open-code-review#1056: add Jinja template support to the review configuration flow.
• desktop/desktop#22767: hide pull-request suggestions when a branch has no commits ahead of the default branch.
• parcel-bundler/parcel#10352: parse UTF-8 BOM JSON correctly in the resolver path.
• Tencent/wujie#1105: avoid the reserved style prop warning in the Vue 2 adapter.
• Tencent/vConsole#737: restore inherited event properties in log output.
• bytedance/flowgram.ai#1176: scaffold projects without partial mutations.
• alibaba/formily#4363: export static guide routes in the docs build.
• pnpm/pnpm.io#908: document confirmModulesPurge.

🧭 Search and coordinates · 1 open PR

• organicmaps/organicmaps#13423: support space-separated DMS coordinates in search.

📊 Full footprint

_{Auto-aggregated from every public PR against external repositories, and refreshed daily by a GitHub Action — dynamic numbers stay current without manual edits.}

Project	⭐	PRs	✅ Merged	🚀 Open
open-city-ai/haidian	406	28	27	0
agentscope-ai/agentscope	30306	2	1	0
alibaba/open-code-review	21764	2	1	1
alibaba/hooks	14977	1	1	0
ant-design/ant-design	99344	1	1	0
Tencent/cherry-markdown	4857	1	1	0
NervJS/taro	37660	6	0	6
bytedance/UI-TARS-desktop	38775	5	0	5
TencentCloud/TencentDB-Agent-Memory	25523	3	0	3
alibaba/formily	12566	1	0	1
assistant-ui/assistant-ui	11966	1	0	0
bytedance/flowgram.ai	8406	1	0	1
desktop/desktop	21811	1	0	1
organicmaps/organicmaps	15271	1	0	1
parcel-bundler/parcel	44023	1	0	1
pnpm/pnpm.io	312	1	0	1
QwenLM/qwen-code	27548	1	0	1
Tencent/vConsole	17519	1	0	1
Tencent/wujie	5025	1	0	1
Total across 19 upstream projects		59	32	24

_{Snapshot 2026-09-01 (UTC) — PRs minus merged minus open are closed without merge and are not counted as adopted work.}

Education

Master's student at BUPT, Electronic Engineering and Information Photonics
Undergraduate background in Network Engineering at BUPT

Contact

dvd@linux.do · z1403594118@gmail.com

📈 GitHub activity

_{Trophy, stats, streak, and calendar cards are regenerated daily by GitHub Actions workflows in this repository.}

_{Animated SVG generated from the public contribution calendar by a scheduled Platane/snk workflow.}