Before building my own stack, I scaled these systems. This is the execution track record behind the projects I’m creating now.
Career Path: From Firmware Engineer to R&D Director
More than twenty years in connected-hardware R&D, starting with surveillance cameras, across three jobs. Each one let me see one layer higher than the last.
Dynacolor (2004–2011): From mechanical engineering into firmware
I studied aerospace engineering, with a mostly mechanical background, and had written almost no code before joining Dynacolor, where I started as a firmware engineer. My first task was focus-motor control for camera lenses. The existing code was written in assembly, and it took about three months of learning before anyone could start developing features on it. I proposed rewriting that part to my manager, and finished it without taking much time.
Around my third year, the surveillance industry was moving from analog cameras to IP cameras. I chose to stay and go through that transition, learning Linux, drivers, and middleware and userspace application development from scratch, and successfully moved the product line from analog cameras to the new trend, IP network cameras. Later, when the Sony platform we used was being discontinued, I led the evaluation and selection of a new SoC platform: sizing it on five- to ten-year shipment volume, negotiating long-term terms with the vendor, and moving the products up to 1080p. That is the stretch at Dynacolor I’m proudest of. When I left, my title was R&D Deputy Manager, owning product lifecycles from kick-off to mass production.
Robert Bosch (2011): Learning how a large enterprise runs R&D through its documents
In 2011 an opportunity came up, and I also wanted to see how a century-old company works, so I joined Bosch as Software & Test Team Leader. At the time, imaging algorithms, application software, and system software were developed mainly in Germany, while hardware R&D was done in Taiwan. I was responsible for about five products, completing three main documents for each, the system architecture document, the system requirements specification, and the test plan, to define its architecture, requirements, and how it would be verified. I also read a large body of internal documents, saw how a large enterprise drives product R&D through documentation, and learned a way of working I carried forward.
Ubiquiti (2011–2026): From team member to leading a 250+ person organization
When I joined Ubiquiti at the end of 2011, the whole Taipei office had fewer than 30 people, and I was one of them, focused on the camera products themselves for the first few years. In 2016–2017 I founded the BSP team and became its lead; the Taipei office then had about 150 people, and the BSP team was around 10. From there I kept building new teams and taking over existing ones, and by 2025 the multidisciplinary organization I led had grown to 250+. A few things I consider most representative of that period:
- BSP as the integration hub: BSP did more than board support. It sat between hardware, firmware, production, and suppliers to communicate and coordinate: when problems came up, nobody passed the buck; we actually solved them and kept the project moving forward.
- Unifying factory test automation: Early on, each product team wrote its factory test programs in its own language and framework, which made maintenance and support hard. I first unified the language (Python, cross-platform, runs on Windows too), then surveyed every internal framework and standardized on the best one, under a two-to-three-year plan. In 2018 I set up the Production Automation team, growing it from a single software engineer. By the end of the plan, about 60–70% of factory testing ran on the unified framework; by 2025 a team of just over ten people supported factory testing for the company’s nearly 100 hardware products a year.
- Taking over the UniFi management software core: We staffed up in Taipei and rebuilt the core functions; it was still running stably after I left.
- The AI data loop: When bringing in AI, I put particular weight on data. The closed loop from data labeling to model training was built from zero (see the AI Core section below).
Most of the way, the work was either zero-to-one or taking over something midstream that needed rebuilding. I didn’t plan a path into management so much as get carried there by how things unfolded. Earning the ICF ACC coaching credential in 2025 helped me most by teaching me to coach myself.
Since 2026: Independent builder
After leaving Ubiquiti in February 2026, I went back to writing code every day. As a one-person company working with AI agents, I ship apps, websites, and embedded tools (see Current Project Milestones below).
14 Years at Ubiquiti: A Legacy of Scaling
At Ubiquiti, my journey was about more than technical delivery; it was about building the human and technical infrastructure that powered a global networking leader.
🚀 Organizational Scaling (0 to 1)
I built and scaled multidisciplinary organizations from scratch to mature, high-performing entities:
- BSP Team: Scaled from 0 to 26+ members.
- Cloud Services: Built the Taipei cloud organization from 0 to 30 engineers.
- Mobile/Apps: Established a unified team of 20+ developers.
- AI Core Development Team: Built from 0 to 15, spanning data engineering, model training, AIOps, AIQA, and edge device delivery — forming a complete AI closed-loop force.
- Factory Automation Software Team: Built from 0 to 10+, focused on intelligent manufacturing software that bridges hardware mass production workflows with digital management, improving yield rates and production efficiency.
- Total Management: Oversaw 250+ professionals across Firmware, Cloud, AI, and QA.
🕸️ Fabric/Matrix: Cross-Functional Synergy
I believe the true strength of an organization lies not in siloed excellence, but in the multiplier effect of cross-team collaboration. I deliberately cultivate a Fabric-style matrix culture — where AI, Firmware, Cloud, QA, and Mobile App teams interweave like the warp and weft of fabric, reinforcing each other precisely when it matters most, and actively preventing the Silo Effect:
- Cross-team knowledge flow: Regular cross-functional Tech Talks and Design Reviews ensure insights flow horizontally across disciplines.
- Shared foundational capabilities: AI inference infrastructure, cloud data platforms, and QA automation frameworks are offered as shared services to all teams, eliminating redundant work.
- Proactively eliminating friction: Regularly mapping cross-team dependencies and integration touchpoints to identify collaboration opportunities and maximize overall effectiveness.
🎯 Coach Leadership: Empower, Don’t Command
As an ICF Associate Certified Coach (ACC), I don’t aim to be the smartest person in the room — I aim to make every team member and every team the autonomous owner of their own domain:
- Trust and delegation: Giving engineers true ownership and accountability over their work without micromanagement.
- Coaching conversations: Asking questions instead of giving answers, guiding team members to discover solutions themselves and build decision-making confidence.
- Minimizing over-intervention: Resisting the urge to step in unless a critical risk demands it — allowing teams to learn through execution and grow through challenge.
- The highest achievement is enabling others: When every team can independently carry a critical role in the organization — even beyond my own capabilities — that is when leadership has fulfilled its purpose.
☁️ Cloud Platform Scale & Growth
Transformed and managed the core Central Cloud Services powering the entire hardware ecosystem globally:
- SSO Identity Platform: Scaled the identity platform to 10 Million registrations (9 Million verified users). Sustained a remarkable 3,400× growth factor since launch, achieving a recent processing peak of ~157k new users/month with a 19% YoY 2025 growth engine.
- IoT Ecosystem Scale (NCA): Expanded global console orchestration to manage 3.5 Million active online consoles (+150% over 3 years) along with 30 Million total managed application devices. Mastercrafted the architectural transition through an explosive 85× early adoption multiplier to sustained 30%+ YoY market maturity.
- High Availability (SRE/SLA): Built robust SRE practices and achieved consistent 99.5% monthly availability across all major cloud services in the past 2 years. Dramatically bound global downtime (limiting critical incident durations to less than 10-12 aggregate hours globally per year out of 8,760 hours).
- Critical Service Expansions: Oversaw the mass-scaling of the Firmware Update Service and active subscription platforms.
🤖 AI Core Development Division: End-to-End Closed-Loop Architecture
We have established a complete AI Closed-Loop Architecture — a self-reinforcing, continuously evolving system that spans from raw data ingestion all the way to edge device delivery. Each stage is tightly integrated, enabling rapid AI iteration and real-world deployment at scale.
Raw Data → Data Labeling → Model Training → AIOps → AIQA → Edge Device Delivery
↑ |
└─────────────────────── Feedback Loop ────────────────────────────────┘
📊 Data (Data Infrastructure)
- Built diverse data ingestion pipelines spanning camera streams, IoT sensor events, and user behavior logs.
- Designed a data warehouse architecture with strong emphasis on data quality, traceability, and compliance (GDPR/privacy standards).
- Introduced data versioning to ensure every model training run maps to a complete and reproducible dataset snapshot.
🏷️ Data Labeling
- Built an in-house labeling platform and workflow covering image classification, object detection (bounding boxes), and semantic segmentation.
- Applied Active Learning strategies to prioritize labeling of the most model-impactful hard samples, reducing labeling costs by 40%+.
- Established QA Audit processes to ensure labeling consistency and accuracy are maintained to a high standard.
🧠 AI Model Training
- Led the full training lifecycle from research prototype to production-ready models, spanning Computer Vision (detection, tracking, recognition) and multi-modal LLM integration.
- Built MLOps training infrastructure supporting distributed training, hyperparameter optimization (HPO), and experiment tracking (MLflow/W&B).
- Achieved model compression optimized for edge hardware via Quantization, Pruning, and Knowledge Distillation — significantly reducing inference latency and power consumption without sacrificing accuracy.
⚙️ AIOps (AI Operations)
- Built a full model lifecycle management platform covering model versioning, A/B deployment, canary releases, and automated rollback.
- Implemented Model Drift Detection to monitor live inference quality and trigger automated retraining pipelines.
- Designed a dual-layer cloud + edge AI inference architecture that dynamically routes workloads based on latency and accuracy requirements.
✅ AIQA (AI Quality Assurance)
- Established a systematic AI evaluation framework covering functional metrics (mAP, Recall, Precision), edge case testing, and adversarial stress tests.
- Defined AI Product Launch Criteria to ensure every AI-enabled product passes rigorous quality gates before shipment.
- Integrated a CI/CD for AI pipeline, embedding model evaluation into the development loop to accelerate iteration velocity.
📦 Edge Device Delivery (Camera & AI Key)
- Led end-to-end development of the UniFi Camera Gen2/3/4/5/6 series, embedding AI capabilities (face recognition, vehicle detection, anomaly analysis) directly on-device for real-time, low-latency inference.
- Drove the development and mass production delivery of the AI Key (dedicated AI accelerator hardware), enabling an AI compute upgrade path for existing devices without hardware replacement.
- Established an AI model OTA update mechanism allowing model-only upgrades independent of system firmware, shortening AI feature iteration cycles to days.
- Managed the complete production lifecycle from RD → DVT → PVT → MP, ensuring shipping quality targets and delivery commitments for AI hardware are met.
🧩 LLM Internal Services: AI-Powered Intelligent Workflows
Beyond edge AI product R&D and delivery, our AI team simultaneously leads multiple LLM core services targeting both internal operations and external user experiences — translating the power of language models into real, deployable business value.
📣 Social Hearing
- Built an automated social media data crawling and analysis pipeline covering major platforms (Reddit, Twitter/X, community forums) for user reviews and feedback.
- Applied LLM-based semantic understanding and sentiment analysis to precisely extract real user pain points regarding features, quality, and service, automatically categorizing them into Actionable Insights.
- Fed analysis results directly into Product Team iteration prioritization decisions, shortening the cycle from “user voice” to “product improvement.”
💥 Device Crash Analysis
- Built an LLM-powered crash log semantic analysis system that automatically parses Crash Reports, Stack Traces, and system logs to identify Root Cause patterns.
- Automatically clusters and aggregates crash events into a ranked hotspot list, enabling Product Teams and firmware engineers to precisely target high-frequency issues and significantly improve bug fix efficiency.
- Integrated with the CI/CD pipeline to automatically monitor crash trends after new releases, enabling early warning detection.
📋 Product Support File Analysis
- Built an automated Support File parsing engine that uses LLM to analyze device support files (logs, config snapshots, status reports) returned by users, extracting key anomaly signals.
- Automatically generates issue summaries and resolution recommendations, helping the Technical Support team quickly pinpoint problems and reducing manual analysis time by 60%+.
- Accumulated analysis results build a knowledge base that empowers the Product Team to precisely understand real user pain points, driving continuous optimization of product design and firmware.
🌐 Company Website & App LLM Backend (Intelligent Support Bot)
- Led the development of the LLM backend services for the company website and mobile apps, providing a one-stop intelligent conversational interface covering product information queries, feature usage guidance, and troubleshooting diagnostics.
- Adopted a RAG (Retrieval-Augmented Generation) architecture, integrating product documentation, knowledge bases, and FAQs to ensure Chat Bot response accuracy and timeliness.
- Designed multi-turn conversation management to support complex scenario context tracking, enabling users to complete configuration, troubleshooting, and service requests without human agent intervention.
- Continuously optimizes models and the knowledge base through user conversation data, forming a positive cycle of “user interaction → knowledge refinement → service quality improvement.”
🚀 Current Project Milestones
Since going independent, I’ve continued to ship:
- OmniTypist v0.6.1 Shipped — macOS AI dictation app with Developer ID signing, notarization, and Sparkle auto-updates fully delivered
- 5 MCP Servers Built — AI-HIL embedded automation (JTAG, serial, vision, audio, build-flash)
- Engram Phase 7 Complete — MCP server + ML worker + Studio UI end-to-end integration
- Post-Quantum VPN Handshake Working — Plexus-Myelin X25519 + ML-KEM-768 hybrid key exchange
- 138 Tests Passing — Plexus-Myelin test suite covering NAT classification, STUN, PQC
- Phygitao Platform Phase 4.3 Hardened — Batch signing, uplink, observability
- 30+ Technical Proposals Written — Covering Engram Client, Voice Fabric, vertical applications (Factory/Farm), SASE architecture, and more