Skip to content
Nat Lee

Keyboard shortcuts

Turn these off if they get in the way of a screen reader or voice control. The setting is saved in this browser.

1
Home
2
Work
3
Projects
4
About
5
Experience
/
Search projects
?
Show this panel

Work

What I built at work

Projects from the companies I worked at. The code stays with them, so each chapter covers what the job needed, my part and what shipped, plus the open-source projects that grew out of it.

Career at a glance

Drawn to calendar scale, so overlaps show. The striped strip is a stretch with nothing on it, squeezed down. Dots are open-source spin-offs and squares are the side projects I pin on GitHub, each placed where it started.

Study
  1. Master's thesis · NTUST
Contract
  1. Contract work
Spin-offs
  1. 2023.04 – PresentHD RenewablesFrom solar forecasts to Japan's power market
  2. 2021.12 – 2023.04Big DataSentiment over three billion posts
  3. 2019.10 – 2021.03InfortrendComputer vision on edge devices
  4. 2018.07 – 2019.10Leopard MobileCrawlers, public opinion and a thesis

2023.04 – Present

HD Renewables

From solar forecasts to Japan's power market

Roles

  1. Senior Engineer2023.04
  2. Senior Engineer2024.12 · Transfer
  3. Principal Engineer2025.09 · Promotion
  4. Deputy Manager2026.03 · Promotion
  5. Deputy Manager2026.05 · Transfer · Star Trade Co., Ltd.
All roles on the Experience page

Meanwhile, on my own time

I joined the Smart Services Division to forecast solar output for an energy-storage demonstration site at a FamilyMart store in Tainan, and to speed up the power-matching algorithms around it.

At the end of 2024 I moved to the International Power Trading Division and built spot-price and spread models for the Japan and Australia trading desks. As Principal Engineer I took technical ownership of that stack; as Deputy Manager I started the data platform that collects Japan's market and weather data. In 2026 I moved to the group's subsidiary Star Trade to lead the engineering of overseas power trading, including three months on site in Tokyo with the vendors.

  • Energy & forecastingAI & machine learning2024.12 – Present

    Japan Power Trading Dashboard

    Spot-price forecasting and trading strategies for the Japanese power market, later extended to Australia. It turns out electricity prices have moods too.

    Results at the time

    • Successfully deployed prediction models for Japan market
    • Achieved 85% accuracy in price trend prediction
    • Reduced trading decision time by 60%
    Read more
  • Energy & forecastingAI & machine learning2023.04 – 2024.12

    Solar Power Prediction System

    Time-series models that forecast solar output for an energy-storage demonstration site, scheduled with Airflow. The sun is punctual; the clouds are the problem. The dashboards here are redrawn on simulated data.

    Results at the time

    • Achieved 92% accuracy in daily predictions
    • Reduced energy storage costs by 25%
    • Successfully deployed to demonstration site
    Read more

Figures are from internal experiments at the time and can’t be checked from outside.

2021.12 – 2023.04

Big Data

Sentiment over three billion posts

Roles

  1. Senior Software Engineer2021.12
All roles on the Experience page

Meanwhile, on my own time

Big Data's products, Keypo and DailyView, turn crawled posts into online buzz and sentiment reports. I led the sentiment engine behind them: the model that labels every crawled post, and the Docker services that run it over the daily crawl stored in Elasticsearch.

Around it I built monitoring dashboards for different clients, internal annotation and data-management tools, and a customer data platform on FastAPI and Elasticsearch that recognises the same user across websites.

  • Language & sentimentAI & machine learning2021.12 – 2023.04

    Sentiment Analysis Platform

    Led the sentiment engine behind a company's social-listening products, reading more than three billion crawled posts. Lesson learned: the internet is mostly complaints.

    Results at the time

    • Successfully deployed core sentiment analysis AI engine for company products
    • Processed over 3 billion data points from web crawlers
    • Achieved 30% improvement in emotion classification accuracy
    Read more

Figures are from internal experiments at the time and can’t be checked from outside.

Open source that grew out of it

The company’s code stays private. Each of these repos took one idea from that time and kept going in public.

  1. Big Data · company code

    2021.12 – 2023.04

  2. GURA33 stars on GitHub

    2019 – 2023

    Sentiment classification was the day job here too; GURA and its datasets went public in 2023.

  3. Sentiment Bundle Benchmark11 stars on GitHub

    2026 – 2026.11 · continues GURA

    GURA's successor: the same question, with today's methods compared on fixed splits.

2019.10 – 2021.03

Infortrend

Computer vision on edge devices

Roles

  1. Image Processing Software Engineer2019.10
All roles on the Experience page

In Infortrend's R&D department I built computer-vision applications, several of them on Nvidia Jetson edge devices: an ad recommendation system that reads the crowd in front of a digital sign, human pose recognition for top-down fisheye cameras in narrow spaces, and heat maps that show retail stores where their products get attention.

On the medical side I worked on MRI and PET brain preprocessing, including automatic AC-PC correction, for an Alzheimer's disease diagnosis project.

  • Vision & imagingAI & machine learning2019.10 – 2021.03

    Persona Tracking

    A camera next to a digital sign tracks the people in front of it, guesses gender and age group, notices who is actually looking, and picks the ad that fits them. Most people, it turns out, walk past.

    Results at the time

    • Shipped an embedded ad recommendation system that reads the demographics of the crowd in front of a screen
    • Later rebuilt as a personal project, Open Persona, with whole-body tracking and interest estimates
    Read more

Figures are from internal experiments at the time and can’t be checked from outside.

Open source that grew out of it

The company’s code stays private. Each of these repos took one idea from that time and kept going in public.

  1. Infortrend · company code

    2019.10 – 2021.03

  2. Simply Blur Detector1010 stars on GitHub

    2019 – 2026.11

    The small CNN that told sharp camera frames from blurry ones, rebuilt with a browser demo.

  3. Fisheye Warping1616 stars on GitHub

    2020 – 2026.11

    The fisheye-to-panorama step that came before pose recognition.

  4. OpenPose Wrist Heatmap33 stars on GitHub

    2020 – 2021

    The motion heat maps from the retail work.

  5. Batch Crop Images7676 stars on GitHub

    2019 – 2026.11

    Crops a whole folder of images with one box, a chore from the image work.

  6. Brain MRI Pipeline4040 stars on GitHub

    2020.01 – 2026.11

    The brain MRI preprocessing from the medical-imaging work: AC-PC alignment, skull stripping and segmentation.

  7. Blur Generator7070 stars on GitHub

    2021.05 – 2026.11

    Makes blurry training data, the other half of the blur detector.

  8. Face Wall Generator11 stars on GitHub

    2021.03 – 2026.11

    Lines up the same wall of faces every time to test face detectors on, a test rig from the vision work.

2018.07 – 2019.10

Leopard Mobile

Crawlers, public opinion and a thesis

Roles

  1. Software and AI Engineer2018.07
All roles on the Experience page

My first job, in Leopard Mobile's Tool Product Division, while I was still writing my master's thesis. I wrote crawlers for Twitter, Facebook and Telegram, kept the public-opinion database in MySQL, and built the NLP behind an ICO sentiment platform, Rating Token & Coin Master. Around it came content classification for a voice assistant, a Taiwanese Hokkien song recognition model for smart speakers, a Chinese reading-comprehension service, and the annotation platforms that fed those models.

There is no single flagship project from this job. What lasted are the tools around it: the crawlers and labeling platforms, the experiments for my thesis, and two contract jobs on the side.

Open source that grew out of it

The company’s code stays private. Each of these repos took one idea from that time and kept going in public.

  1. Leopard Mobile · company code

    2018.07 – 2019.10

  2. MyAnimeList Review Crawler44 stars on GitHub

    2019 – 2026.11

    A review crawler from the public-opinion work, used to collect sentiment data.

  3. Word Net11 stars on GitHub

    2018 – 2026.11

    Started as a way to collect how happy or sad people find a sentence; now it is a game.

  4. Reading Comprehension Labeling33 stars on GitHub

    2019.07 – 2026.11

    The labeling tool for the reading-comprehension data. I did the labeling myself too.

  5. Multimodal Labeling22 stars on GitHub

    2019.05 – 2026.11

    The same idea for text-classification labels.

  1. Master's thesis · NTUST

    2017 – 2019

  2. IMDb Preprocessing Study11 stars on GitHub

    2019 – 2026.11

    Thesis research: re-running a paper's IMDb experiments on how text preprocessing changes sentiment models.

  3. MyAnimeList Review Crawler44 stars on GitHub

    2019 – 2026.11

    Built the MyAnimeList review dataset (MALCoD) during the thesis work.

  4. GURA33 stars on GitHub

    2019 – 2023

    The 2019 BiGRU + attention model written to introduce those datasets.

  5. Sentiment Bundle Benchmark11 stars on GitHub

    2026 – 2026.11 · continues GURA

    GURA’s successor: the same datasets, today’s methods, fixed splits.

  1. Contract work

    about 2018 – 2019

    A blockchain project for a financial training institute and a XOOPS website for a school. Only the blockchain work has a public repo.

  2. Blockchain from Scratch11 stars on GitHub

    2018 – 2026.11

    Grew out of the blockchain contract job.

Roles, education and skills are on the Experience page. Things I built on my own time are under Projects.