Building AIoT and Edge AI Products That Work in the Real World
I build intelligent systems across smart home, edge devices, private LLM deployment and fine-tuning, Java backend infrastructure, and real-world AI product engineering.
Smart-home AI host, Jetson edge vision, private LLM deployment and fine-tuning, gateway and control panels.
From Java distributed systems to smart-home AIoT.
Luke is a hands-on engineering builder focused on real-world delivery — with experience across smart-home AIoT, edge devices, on-prem AI hosts, private LLM deployment and fine-tuning, smart-home gateway development, Java backend, distributed systems, and fintech.
Today the work spans smart-home AI hosts (Home Assistant + HomeClaw), Jetson edge vision and audio-event recognition, private LLM deployment and fine-tuning across server and device (vLLM, RKLLM), tri-mode PLC + WiFi + BLE 5.2 hub gateways, Android control panels, and the Hommor AI host lineup.
The throughline: turn AI capability into something a real device, in a real home, can run reliably — not a demo.
Today’s work, across six fronts.
Six adjacent fronts. They add up to one thing: AI running on devices, in the home.
Smart Home AI
01 / 06A local AI host that drives the home.
Combine local AI hosts, camera streams, voice interaction, and device orchestration into one smart-home stack — so household automation actually feels intelligent.
Local InferenceCamera StreamVoiceAutomationEdge Vision Agent
02 / 06On-device vision and audio-event recognition.
Run vision on Jetson-class edge devices to recognize people, actions, gestures, pets and falls; run audio-event recognition for baby cries, doorbells, knocks, smoke alarms and more — all on-device, with low latency, pushing into smart-home flows.
Jetson Orin NanoMiMo-VLAudio EventsCamera StreamScene TriggerPrivate LLM: Deploy & Fine-tune
03 / 06OpenAI-compatible LLM service across GPU and NPU, plus fine-tuning.
Practical private-LLM serving across server and device — vLLM on DGX Spark and 4×RTX 4090, plus Qwen3-0.6B on RK3576 NPU via RKLLM. Unified OpenAI-compatible API. Qwen-series fine-tuning targeted at the smart-home domain.
vLLMRKLLMQwen3Gemma 4RK3576 NPUOpenAI-compatibleAI Voice & Knowledge Base
04 / 06Voice in, knowledge out, devices respond.
Exploring AI voice interaction, enterprise knowledge bases, Dify, the Qwen series, and local inference / fine-tuning in smart-home scenarios.
VoiceRAGDifyQwenFine-tuningGateway & Control Panel
05 / 06The device-side engineering of an AIoT platform.
Build the device side around tri-mode PLC + WiFi + BLE 5.2 hub gateways, ZigBee gateways and Android control panels, with local-LAN control, OTA, and log upload — engineered as one fleet, not one-off boards.
PLCWiFi+BLE 5.2ZigBeeAndroid PanelOTALocal ControlAI Coding Tools
06 / 06AI-native dev tooling for private networks.
AI coding tools for private enterprise dev environments and intranet code scenarios — pushing engineering velocity in places the cloud cannot reach.
AI CodingIDEPrivate ModelEnterprise
Projects across device, edge, model, and platform.
Eight projects spanning home AI hosts, edge devices, private LLMs, distributed backends, and financial systems.
Smart Home AI Host
A local AI host that puts the home into WeChat and Feishu — no extra app to install.
Smart-home shouldn’t mean another app for the family. This local AI host pulls every device into one place — Home Assistant for the ecosystem, an in-house HomeClaw bridge to the on-prem hub gateway — and surfaces control inside the chat tools the family already uses every day.
- In-house HomeClaw (built on OpenClaw) bridges Home Assistant and the hub gateway
- Control entry: WeChat + Feishu (tools the family already uses)
- Local service stack — DPanel, Frpc — orchestrated with Docker
Jetson Vision Agent
Local vision and audio-event recognition on a Jetson Orin Nano.
A single Jetson Orin Nano NX Super 8G handles the home’s eyes and ears on-device — video and audio never leave the box, and one board runs both vision understanding and audio-event classification at low latency.
- Vision model: MiMo-VL / Miloco 7B; audio: event-classifier head
- Live preview, 5-minute rolling clips, time-aligned playback
- Vision events — people, activity, gestures, pets, falls
- Audio events — baby cries, coughs, snoring, doorbells, knocks, pets, smoke alarms
Private LLM Stack
One LLM stack from 4×RTX 4090 down to RK3576 NPU.
Same LLM stack, three very different hardware targets — server racks, a desktop AI supercomputer, and an edge NPU. Each runs the model that fits, all behind a single OpenAI-compatible API.
- Server: 4×RTX 4090 + Tensor Parallel — Gemma 4 31B & Qwen3.5 27B, GPU-exclusive switching
- Desktop AI host: DGX Spark (ARM64 / GB10 / CUDA 13) running vLLM + Gemma 4 31B
- Edge NPU: RK3576 + RKLLM running Qwen3-0.6B for on-device intent parsing
- OpenAI Responses + Conversations API, API-key auth
Koder
A Chinese AI coding IDE for enterprise intranet R&D.
A Chinese AI coding IDE for enterprise intranet R&D — based on the Void AI editor, with the AI panel, settings, and onboarding rewritten in Chinese at the source level so the experience feels native, not translated.
- Core: Void AI editor (VS Code 1.99.x base), full re-branded fork
- Source-level localization: Chat → 对话, Gather → 收集, Agent → 智能体
- Bundled MS-CEINTL language pack, zh-cn default on first launch
AIoT Gateway & Control Panel
The device side of a smart-home SaaS / PaaS — modules to gateways to panels.
End-to-end engineering of the device side of a smart-home SaaS / PaaS — modules, gateways, control panels, on-device protocols, and remote ops — running across the Hommor AI host lineup from RK3566 to Jetson Orin Nano and Kunpeng 920.
- Gateways: tri-mode hub (PLC + WiFi + BLE 5.2, SSD202D + ESP32-C6, IEEE 1901.1, up to 500 nodes); ZigBee / PLC / WiFi+BLE fleet
- Terminals: Android in-wall control panels and the ESP32-S3 “Xiao Bai” voice box (custom wake-word + MultiNet7 Chinese model)
- Comms: UDP self-discovery + local-LAN MQTT, parallel to cloud MQTT
- Ops: SN-hashed subdomain tunnel, incremental OTA, sub-device batch upgrade
Distributed Backend Infrastructure
The platform and infra other teams build on top of.
Long-running ownership of core services and platform infrastructure — core services (API gateway, identity, RBAC, payments, scheduled jobs) and the platform layer (MySQL, Redis, Nacos, RocketMQ, Elasticsearch, Kibana, SkyWalking, Prometheus, Grafana, Jenkins, GitLab, YApi, OpenLDAP, OpenVPN, JumpServer) — plus backend engineering standards.
- Java distributed systems
- Microservice governance
- Core middleware build-out and operation
- Monitoring, alerting, observability
- CI / CD pipeline (Jenkins, GitLab)
- Backend engineering standards
ABS Securitization Platform
Asset-backed securities — from base assets to issuance and lifecycle management.
A securitization platform built at Didi for the group and ecosystem partners. Pool owned assets, isolate risk, enhance credit, and issue securities backed by future cash flows on public or private markets — lowering financing cost. Spans the full ABS workflow: asset data management, product design, issuance, and ongoing lifecycle management.
- Tech lead: 0 → 1 system delivery, first-version ABS shipped
- Issued asset-backed special-purpose plans (ABS)
- Base-asset onboarding and data management
- Risk isolation, credit enhancement, cash-flow modeling
- Lifecycle management for the duration of the security
Lending, Payment & Settlement Systems
Lending, payment, and settlement — money-moving systems built for reliability.
Lending, payment and settlement systems shipped at Qudian and Da Shu Tian Kong — covering trading, insurance, membership, pre-loan risk, post-loan operations, and settlement payouts. Integrated end-to-end with funding-source partners (JD Tech, JD Finance, Xinwang Bank, Taikang Insurance, Jinmeixin) for credit, contract signing, statement sync, and repayment.
- Trading: borrow / repay / deduct / refund / acquiring
- Insurance: quote, underwrite, payment confirmation, refund
- Pre-loan risk: credit lines and contract signing with funding sources
- Post-loan: external funding-source integration, SFTP delivery, statement sync, repayment
- Settlement: payouts via Alipay and bank cards, freight insurance refunds
- Membership: balances, withdrawals, reward distribution
A decade of building load-bearing systems.
From financial business systems, to platform infrastructure, to today’s AIoT and AI engineering.
- 2025.05 — PresentHardware & Model Tech Lead
Jin Yun Zhi Lian
Leading technical build-out across smart-home AI hosts, edge inference, video recognition, private LLM deployment and fine-tuning, smart-home gateways, control panels, and AI-voice ↔ device integration.
- 2021.11 — 2025.04Staff Engineer
Kingsoft Cloud
Owned core services, platform infrastructure, and the AIoT platform — covering API gateway, identity, RBAC, payments, scheduling, middleware, monitoring, CI/CD, and the smart-home platform.
- 2020.05 — 2021.11Senior Engineer
Didi
Built securitization (ABS) and interbank-flow platforms — 0 → 1 system design, core development, and production delivery in financial domains.
- 2019.05 — 2020.05Senior Engineer
Da Shu Tian Kong
Worked on business and post-loan systems — external funding-source integrations, file parsing, SFTP delivery, statement sync, and repayment.
- 2016.11 — 2019.05Senior Engineer
Qudian
Worked across trading, insurance, membership, pre-loan risk, and settlement — payments, deductions, policies, credit lines, contracts, and settlement payouts.
Tech Stack.
AI engineering and edge devices are today’s focus; backend, data, infrastructure, and monitoring are the foundation.
AI Engineering
Edge & Device
Backend
Data & Middleware
Infrastructure
Monitoring & Ops
Let’s build something real
Best reached over email — for project work, AIoT collaboration, or a technical conversation, drop a line.