Open Source Project

Welcome to my website!

I’m Zephyr, an AI engineer based in Taiwan.

What's Here

The page you’re currently viewing is designed for technical documentation.

• For resources on paper readings, head over to Papers.
• My blog can be found at Blog.

＊

I’ve completed several projects on GitHub, including:

1. AutoTraderX: This is a record of my experience integrating systems with Taiwanese stock brokers.
2. DocsaidKit: My personal toolbox.
3. GmailSummary: A project where I experimented with integrating Gmail and OpenAI.
4. WordCanvas: A tool I developed to synthesize text images, primarily for creating training data.
5. 【Deep Learning Project】DocAligner: A project designed to locate the four corners of a document.
6. 【Deep Learning Project】DocClassifier: A project focused on document similarity matching.
7. 【Deep Learning Project】MRZScanner: A project designed to recognize the MRZ area on a document.

Most of my projects are private, some are still under development, and others are complete but need further optimization.

Are You an Engineer?

Yes.

I enjoy solving real-world problems using scientific knowledge, and I find great joy in doing so.

＊

A few years back, those working with images might have called themselves “Computer Vision Engineers,” those working with text might have been “Natural Language Processing Engineers,” and there were also “Machine Learning Engineers” and “Deep Learning Engineers.” Within certain circles, these titles even created a hierarchy.

But regardless of the specialty, everyone looked down on the title “AI Engineer”:

• "Oh, another company/job seeker trying to claim a flashy title without real skills!"

However, times have changed. We’ve witnessed the unification of text, images, audio, and other types of data into a high-dimensional space under the wave of “Foundation Models,” which have become essential for breakthroughs and research publications.

People have since realized that these engineers are essentially doing the same work, just in different dimensions.

Soon, the distinctions between these titles faded, as cross-disciplinary research became necessary. Researchers now need to understand multiple fields to advance their work. This shift has made it difficult to clearly define what we do.

And so, we’ve come full circle:

• Yes, that’s it! AI Engineer!

So What Is This Place?

Not long ago, I stumbled upon Meta’s Docusaurus while browsing, and I was impressed with its capabilities.

I thought, why not use it to build a blog? So I combined Docusaurus with GitHub Pages for automatic deployment, and now you’re looking at the result.

info

Surprise! This means if GitHub ever goes down, so will this site.

So, let’s take good care of GitHub and hope it never breaks. (What a conclusion!)

---

I believe the hardest part of “creating a website” is naming it.

Analyzing text is my daily job—whether it’s OCR, image fraud detection, topic classification, keyword extraction, or similar tasks. From my perspective, text isn’t limited to just words; it could be an image, a video, a song, a dataset, or even human behavior. As long as it has analytical value, or we’re willing to analyze it, anything can be considered text (chaotic neutral?).

So, in the end, I wanted the name of this site to be related to this field, and I chose DOCSAID.

This name is a blend of two words: “DOC” and “SAID,” roughly meaning:

• The moment a text is created, it has already conveyed its message.

So, what exactly did these texts say? That’s what we need to analyze!

Interestingly, after picking this name, I realized it also contains the letters AI—an unexpected delight.

＊

Feel free to click around the menu on the left. I’ve already completed some sections.

If you find empty content, it means I’m still working on it, so please be patient.

How to Fine-Tune Models

This might be the topic you’re most interested in.

Based on the topics I’ve defined and the models I’ve provided, I believe they can solve most application scenarios.

However, I understand that some scenarios might require better model performance, which means you’ll need to collect your own dataset and fine-tune the model.

If you find yourself stuck at this step, don’t worry—you’re not alone. Most people struggle here.

• Scenario 1: You know my project meets your needs, but you don't know how to fine-tune it

  In this case, you can email me with your requirements and provide your dataset. I can help you fine-tune the model, giving you better results.

  I don’t charge for this, but there’s no deadline, and I can’t guarantee I’ll do it (this is important!).

  Although I work on open-source projects, I’m not doing this full-time. If the timing is right, I’ll update the model, and you might get a better result just by sending an email. It’s a win-win, right?

• Scenario 2: You want to develop a specific feature

  Email me, and we can discuss it. If I find it interesting, I’d be happy to help develop it, but I’d like you to prepare a dataset of a certain scale first. Even if I’m interested, I might not have the time or access to the necessary data, especially if it requires specific channels.

  As with the first scenario, I won’t charge, but there’s no deadline, and I can’t guarantee it’ll happen.

  tip

  If the feature you want is for a public competition? The answer is no. Competitions often have copyright and other restrictions, and I don’t want to get into trouble with the organizers.

• Scenario 3: You need to quickly develop a specific feature

  When time is of the essence, we can shift to a commissioned development approach. Based on your requirements, I’ll provide a reasonable quote.

  Additionally, the ownership of the commissioned project must be discussed upfront. Generally, I retain ownership, and you can freely use it. I don’t recommend buying out the project, as it goes against the principle of continuous improvement. With technological advances, today’s solution may soon be replaced by a newer method. If you buy out a project, you might find that your investment loses value over time.

  tip

  You might not fully understand the issue of project ownership.

  Think about it carefully—you might just want to “drink the milk” without having to “own the cow.”

  • Owning a cow is a lot of work. (You’d need engineers to maintain the project)
  • It takes up space and is hard to care for. (You’d need to set up training machines, and cloud servers are expensive, while physical servers can break down)
  • It’s sensitive to the environment. (Hyperparameter tuning can drive you crazy)
  • And it could just die unexpectedly. (Results might not meet expectations)
  • It’s a real hassle. (Spending money to buy out a project)

  Also, the most valuable part of most projects is the dataset, followed by the thought process behind the solution.

  Without open-sourcing a private dataset, having access to the code is mostly just for viewing purposes.

  If, after careful consideration, you still insist on buying out a project, I won’t stop you. Let’s do it.

  info

  • In all forms of development projects, we will never open-source the data you provide unless you give explicit permission.
  • Under normal circumstances, the data will only be used for model updates.
  • Dataset submission: docsaidlab@gmail.com

Finally 🍹

I have invested a lot of time and effort in these open-source projects. On the one hand, I want to improve my technical skills, and on the other hand, I hope to make some contributions to the community.

I believe that every small contribution is a part of our collective effort to create a better world.

If you find my open-source projects helpful or appreciate my efforts, you can support me through “Buy Me A Coffee.” This will help me continue to maintain and develop these projects. In addition, feel free to leave a few words or give a star to the project. These are all great encouragements to me.

Every bit of support, no matter how small, motivates me to continue improving existing projects and developing more interesting new tools.

＊

2024 Zephyr