Stanley.
AI Software Engineer at Garmin Ltd., building the company’s GenAI platform and real-time computer vision systems. I serve and evaluate LLMs/VLMs, build tool-calling AI agents, and ship AI from research into production.
Hi — I'm Kuan-Wei Lu, but call me
AI Software Engineer
Garmin Ltd.
GenAI Platforms
Real-time Computer Vision
About Me
I am an AI Software Engineer II at Garmin Ltd. (Nov 2023 – present, promoted Jun 2026), where I own the company’s enterprise GenAI platform — LLM/VLM serving on vLLM with an automated evaluation framework — now scaling to 30+ production use cases across 8 teams and 2M+ requests/week.
I also build and operate a real-time computer vision platform across 5 sites, processing 63M+ telemetry events/month to deliver critical operational alerts. Most recently I am building a multimodal tool-calling AI agent on LangGraph that orchestrates detection, OCR, and speech workflows from natural language. I mentor engineering interns and serve as a technical interviewer.
Previously, I completed my Master’s in Electrical Engineering at NTHU, conducting computer vision research under Prof. Min Sun, which resulted in a publication at ICRA 2021.
Technologies I work with:- Python
- LLM / VLM Serving
- LangGraph / LangChain
- Computer Vision
- Docker / Kubernetes
- Kong API Gateway
- Apache Kafka
- Grafana
- vLLM
- LLM Evaluation (Ragas)
- PyTorch / OpenCV
- MySQL / MongoDB
- C++ (CUDA/OpenMP/MPI)
Experience
Jun 2026 – present
- Promoted to Engineer II after 2.5 years owning the organization’s GenAI and computer vision platforms.
- Building a multimodal tool-calling AI agent on LangGraph and an internally hosted LLM that orchestrates multi-step workflows — detection, OCR, TTS/ASR — from natural language prompts.
- Continuing to raise the team’s engineering bar as a code reviewer and technical interviewer, and mentoring engineers on production AI system design.
AI Software Engineer I
Nov 2023 – Jun 2026
- Built and scaled the enterprise GenAI platform — LLM/VLM serving on vLLM, an automated evaluation framework, and a Kong API gateway — serving 30+ use cases across 8 teams and 2M+ requests/week.
- Led company-wide GenAI adoption by evaluating 10+ LLMs/VLMs through structured POCs against internal benchmarks, producing technical assessments that guided strategic model selection and shipped 5 core production use cases.
- Served as a code reviewer for the AI team and a technical interviewer for 1+ year while still at this level, contributing to 2 full-time hires.
- Mentored engineering interns on AI system design and production deployment, guiding them to independently ship production-grade services.
Projects
Multimodal Tool-calling AI Agent
LangGraph
LangChain
LLM Agents
Multimodal
End-to-end Owner at Garmin (May 2026 – present).
Building a stateful tool-calling agent on LangGraph with an internally hosted LLM as its reasoning core, orchestrating multi-step workflows — object detection, OCR, and TTS/ASR — from natural language prompts. Designing structured tool schemas and deterministic error-recovery graphs to keep the agent dependable across long, multi-step tasks.
Real-time CV Surveillance & Analytics Platform
Python
Computer Vision
VLM
Apache Kafka
MySQL
MongoDB
End-to-end Owner & Maintainer at Garmin (Nov 2023 – present).
Built an enterprise computer vision pipeline that runs across 5 sites, processing 63M+ telemetry events/month through Apache Kafka and fusing detection, tracking, and pose estimation to deliver ~4,000 critical alerts/month, with results persisted to MySQL/MongoDB for analytics. A major refactor cut per-model GPU memory by 70% (2 GB → 600 MB) and tripled the models served per GPU by deduplicating model instances across containers and adopting FP16 half-precision.
AI-Powered Document Inspection System
Python
VLM
OCR
Computer Vision
Automation
End-to-end Owner at Garmin (Mar 2024 – present).
Built an automated document-verification pipeline that combined rule-based parsing with VLM-based data extraction, saving 5,300+ manual hours a year across 3 business units. Replacing legacy OCR with robust VLM extraction resolved systemic text and barcode misinterpretation, enabling automated compliance validation across hundreds of pages per document.
Enterprise GenAI Platform — LLM/VLM Serving & Evaluation
vLLM
LLM / VLM
Ragas
Kubernetes
Kong Gateway
Grafana
Platform Owner at Garmin (Apr 2024 – present).
Built the company’s automated LLM/VLM evaluation framework (LLM-as-judge, Ragas) and tuned the serving stack — a 2nd-generation VLM tripled inference speed (3×) with higher accuracy. The platform runs on Kubernetes and scales to 30+ production use cases across 8 teams, handling 2M+ requests/week. I designed its API gateway layer on Kong across 16 APIs / 52 projects, sustaining ~99.9% availability with zero production outages, and built Grafana dashboards that surfaced anomalies — Kong blocked ~4,000 unauthorized requests. I also shipped reusable tool-calling configs and skill packages that expose platform APIs to internal AI coding agents.
Robust 360-8PA (ICRA 2021)
Computer Vision
Python
Research
Redesigning the normalized 8-point algorithm for 360° FoV images.
Published at IEEE International Conference on Robotics and Automation (ICRA), 2021.
Education
Jul 2020 – Jul 2022
Master's Degree in Electrical Engineering
National Tsing Hua University (NTHU), Hsinchu, Taiwan
Advised by Prof. Min Sun at the Vision Science Lab (VSLab). Research on semi-supervised learning for 2D object detection.
Resulting in a publication at IEEE ICRA 2021:
Sep 2016 – Jun 2020
Bachelor's Degree in Bio-Industrial Mechatronics Engineering
National Chung Hsing University (NCHU), Taichung, Taiwan
Undergraduate researcher at the Precision & Micro Mechatronics Laboratory (PMML), advised by Prof. Yao-Chuan Tsai (Jul 2018 – Jun 2020).
Get in Touch
Open to new opportunities, research collaborations, and technical conversations. Drop me a line.