I take models off the
research bench and make
them run on the hardware.

01 Selected work

Five things that show what I actually ship, rather than what I can list.

NUCTECH (Tsinghua Tongfang) · Aviation security screening · 2018–2020

A keypoint model small enough for the scanner

Millimetre-wave and terahertz body scanners at civil airports have to flag concealed items between one passenger and the next. The human-keypoint model that fed that pipeline started at over 70 MB — too large for the deployment target.

I rebuilt it around MobileNet backbones with squeeze-and-excitation modules and quantisation, bringing it to 2.6 MB while improving accuracy and holding millisecond-level inference. Separately, I reformulated keypoint localisation as a regression problem, which cut localisation error substantially and removed the need to maintain separate gender-specific models.

Expand

Beijing Institute of Radio Metrology and Measurement · 2022–2023

Perimeter intrusion, judged from posture

A bounding box tells you a person is near a fence. It doesn’t tell you whether they are walking past it or going over it. I built the intrusion alert on human keypoint estimation instead, so the decision rests on limb geometry — the pose of someone scaling a railing is separable from the pose of someone standing beside one.

The system localises the event, captures the frame and dispatches the alarm automatically, replacing a person watching a wall of CCTV feeds.

Expand

Beijing Institute of Radio Metrology and Measurement · 2022–2023

Crowd gathering as a density problem

Counting people is easy. Deciding that a crowd has formed is a spatial question — it depends on local density, not headcount. Detections are clustered per-region and scored against a threshold tuned on site, so a queue at a bus stop doesn’t trigger and a genuine gathering does.

This ran as part of an end-to-end worker-safety stack on Nvidia Jetson with DeepStream: fall detection, PPE compliance (hard hat, hi-vis), smoking detection and virtual geofencing. I owned it from data collection and annotation strategy through training, on-device optimisation, field deployment and false-positive tuning under real site conditions — lighting, occlusion, camera angle, weather.

30 s · screen recording · Chinese UI

Founder & Technical Lead · Beijing Houzhen Visual Design · 2023–2026

File IO — files, handled in plain language

A WeChat mini-program where you describe what you want done to a file and an LLM agent system does it. In the clip, the user uploads a photograph of a room and types, in Chinese, “replace the blue sofa in that PNG with a vintage brown leather one” — the agent selects the right tool, performs the edit and returns a new file.

I built and ran the whole thing alone: Vue.js / uni-app front end, Node.js API, deployed on self-managed Proxmox VE bare metal. Rebuilding the assistant from a single agent to a multi-agent architecture fixed tool-selection degradation and hallucination at scale, and cut token consumption per task by roughly 80%. A RAG pipeline over a 100,000-word bilingual knowledge base (LlamaIndex, Elasticsearch as vector store) is exposed to the agent layer as a first-class tool; Kafka and Redis carry event-stream tracing and response caching, with ELK observability over agent execution traces.

Expand Expand

Personal engineering · ongoing

A production estate I run end to end

Proxmox VE on bare metal, services in LXC containers, behind an OpenWrt gateway with dynamic DNS. My residential ISP blocks ports 80 and 443, so public traffic enters through a minimal cloud VPS and is tunnelled back to the gateway; Nginx Proxy Manager centralises reverse proxying and TLS, and a WireGuard overlay carries routed access between Ireland and China. The same hybrid-ingress pattern cut hosting cost for my company’s services to a fraction of an equivalent all-cloud deployment.

On top of it: Elasticsearch as event store and security telemetry (above), paperless-ngx with bilingual OCR, Trilium, AFFiNE, Memos, Kavita and Jellyfin — and a self-hosted mail server that stays deliverable, which is its own discipline.

03 Experience

1. Nov 2023 – Jan 2026

   Founder & Technical Lead Beijing Houzhen Visual Design Co., Ltd.

   Bootstrapped a product venture around File IO, an LLM-agent-backed file assistant. Product, architecture, full stack and infrastructure, solo. Benchmarked LoRA vs. P-Tuning v2 on ChatGLM2 and Llama 2; fine-tuned Whisper-large-v2 with LoRA for domain speech.

2. May 2022 – Nov 2023

   Algorithm Engineer Beijing Institute of Radio Metrology and Measurement

   Real-time construction-site worker safety on Jetson + DeepStream — falls, PPE compliance, smoking, geofencing, pose-based perimeter intrusion. Also authored the technical proposals that turned this capability into public-sector smart-city tenders.

3. Mar 2020 – Nov 2021

   Algorithm Engineer Lalamove (Shenzhen Easy Van) · Platform risk control

   Multimodal fraud detection across vision and speech. Built an ALPR system from detection + OCR that outperformed Tencent, Alibaba and Megvii on the internal benchmark; a face-embedding search index over a large identity database; and edge detection models for in-vehicle IoT devices.

4. May 2018 – Feb 2020

   Computer Vision Engineer NUCTECH (Tsinghua Tongfang) · Aviation security

   One-stage detection for prohibited items on millimetre-wave and terahertz scanners at civil airports. Model compression from 70+ MB to 2.6 MB; keypoint localisation reformulated as regression.

5. Jul 2016 – May 2018

   Computer Vision Engineer Beijing Institute of New Technology Application

   Palmprint recognition on low-power embedded devices in C++ — hand-designed Gabor filter banks, bitwise matching. Restructured and parallelised the Gaussian filter pipeline with ARM NEON SIMD for a material latency reduction.