Xin Fu

Managing windows effectively with AeroSpace

Wed, 11 Feb 2026 00:00:00 +0000

Managing application windows has been a never-ending battle, especially with AI tools like Claude Code, Codex, and Cursor encouraging work across multiple projects simultaneously. I’ve relied on macOS built-in Spaces , but as the number of workspaces grows, switching and finding the right window becomes harder.

AeroSpace is a tiling window manager for macOS. I tried a similar tool called yabai before but wasn’t quite impressed.

AeroSpace works pretty much out of the box — windows snap into place without overlapping. It implements virtual workspaces in a faster and more predictable way than native Spaces. For example, workspaces are mapped to shortcuts like Option + 1 or Option + A, making switching between them nearly instant.

Television: a modern fuzzy finder for the terminal

Fri, 12 Dec 2025 00:00:00 +0000

I discovered television (tv), a terminal fuzzy finder UI written in Rust, thanks to the blog post Hidden Gems: Part 2 by Zed.

It’s actually very similar to fzf , a command line fuzzy finder which I use mostly for shell integration like Ctrl+R searching for history.

The tv command line provides a convenient way to chain command line tools such as fd, bat, gh, etc. into a fuzzy finder with telescope style layout. It also has shell integrations.

Use custom LLM providers in Claude Code

Mon, 24 Nov 2025 00:00:00 +0000

Claude Code has been one of the most popular coding agent this year. I’ve relied on it heavily for common development tasks. One useful but lesser-known feature is that Claude Code allows you to override the default Anthropic API endpoint using the ANTHROPIC_BASE_URL environment variable.

This is very useful in many scenarios, such as working around quota limits, connecting to custom model endpoints or third-party providers like GitHub Copilot .

Set Google search AI mode as browser default

Sun, 23 Nov 2025 00:00:00 +0000

With the recent announcement of Gemini 3 , I was thinking about giving Google search AI mode another shot.

I’ve been using Perlexitity for a while and found it quite useful for many of my searches, but lately it has felt slower and the quality of results has declined.

After trying out the latest Google search AI mode, I’m quite impressed by the quality, speed and overall user experience. I’ve now set it as my default search option in Chrome.

Run Llama 2 locally on MacBook

Sun, 23 Jul 2023 00:00:00 +0000

Last week, Meta released Llama 2 , an “open source” large language model that is free for research and commercial use. Within a few hours, the community has ported Llama 2 to llama.cpp which makes it eaiser and more efficient to run Llama 2 locally.

Download and compile llama.cpp

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp && LLAMA_METAL=1 make

Note that LLAMA_METAL is set to 1 to enable using GPU on Apple Silicone. On my M1 Pro MacBook Pro, the compliation took about a few seconds.

Create interactive utility app with Streamlit

Mon, 10 Jul 2023 00:00:00 +0000

Streamlit is an open-source Python library that makes it easy to create and share beautiful, custom web apps for machine learning and data science.

Besides data science and machine learning, I found Streamlit can also be used for creating very simple utility apps.

Postboy - a simple Postman-like app

For example, we can create a simple Postman -like app for testing REST APIs with Streamlit in just a few lines of code.

Handling JWT in Python

Sun, 09 Jul 2023 00:00:00 +0000

JSON Web Tokens (JWT) is an open standard (RFC 7519 ) that defines a compact and self-contained way for securely transmitting information between parties as a JSON object.

JWT Structure

JWTs consist of three parts separated by dots:

Header
Payload
Signature

The header contains the algorithm used to sign the token and the type of token. The header is base64 encoded and looks like this:

Question answering over documents with LLM

Sat, 01 Jul 2023 00:00:00 +0000

One of the most popular applications for large language model (LLM) is question answering over various types of documents, such a plain text, web pages, and PDFs. Usually, we want to make the model answer the question which it hasn’t been trained on.

Overview

There are mainly two steps involved:

Data ingestion: load source documents and convert them into vector embeddings which will be stored in a vector database
Question answering: when given input question, convert to vector embedding first, then perform similarity search within the vector database, and top k results will be used as context for the LLM to generate answer to the question.

Diagrams:

Deploy Ghost instance to Fly.io

Wed, 28 Jun 2023 00:00:00 +0000

Disclaimer: Not affiliated with Ghost or Fly.io, this is only for evaluation purpose.

Fly.io comes with generous free tier allowance, which can be used to run some random side projects for free. Ghost is a WordPress like publishing platform. Unfortunately, there’s no free tier provided by their company. Similar to WordPress, it provides an open-source version that can be self-hosted.

Get Started

To get started, first create an account on Fly.io , and install flyctl :

Custom middleware for FastAPI application

Sun, 25 Jun 2023 00:00:00 +0000

FastAPI is a Python web framework for building APIs. By using a middleware, we are able to process request and response before/after they get handled by the application.

Implement custom middleware

Starlette is a lightweight ASGI framework/toolkit on which FastAPI is based. It provides BaseHTTPMiddleware class for us to implement custom middleware. It’s required to override the async def dispatch(request, call_next) method.

from starlette.middleware.base import BaseHTTPMiddleware

class CustomHeaderMiddleware(BaseHTTPMiddleware):
 def __init__(self, app, header_value='Example'):
 super().__init__(app)
 self.header_value = header_value

 async def dispatch(self, request, call_next):
 response = await call_next(request)
 response.headers['Custom'] = self.header_value
 return response

The example middleware above simply adds a Custom header to the response.

Notes on "Langchain for LLM Application Development"

Sun, 25 Jun 2023 00:00:00 +0000

Course Link: LangChain for LLM Application Development - DeepLearning.AI

Previous notes:

Notes on “Building Systems with the ChatGPT API”

Introduction

What is Langchain

Open-source development framework for LLM applications
Provide both Python and JavaScript/Typescript packages

Modular components which can be combined to build end-to-end applications.

Models, Prompts and Output Parsers

OpenAI API

Example:

import openai

def get_completion(prompt, model="gpt-3.5-turbo"):
 messages = [{"role": "user", "content": prompt}]
 response = openai.ChatCompletion.create(
 model=model,
 messages=messages,
 temperature=0,
 )
 return response.choices[0].message["content"]

LangChain

from langchain.chat_models import ChatOpenAI
chat = ChatOpenAI(temperature=0.0)

Prompt template

A prompt template refers to a reproducible way to generate a prompt.

Set up Hugo with Tailwind CSS in 2023

Sat, 24 Jun 2023 00:00:00 +0000

With the release of v0.112.0 , Hugo added the native support for TailwindCSS v3.x . The author of Hugo provided an example repository setting up TailwindCSS v3: bep/hugo-starter-tailwind-basic .

Note that it uses PostCSS so make sure to use Hugo extended version greater than v0.112.0.

How does it work?

According to the release note:

The basic concept is to add hugo_stats.json to the server watcher list in Hugo and trigger a new TailwindCSS build only whenever either this file or the main CSS file changes.

Write pytest tests for argparse

Thu, 22 Jun 2023 00:00:00 +0000

Writing unit tests for Python code that uses argparse can be non-trivial.

Call `main` with optional arguments

One way suggested by Simon Willison was to make main() function take optional arguments:

parser = argparse.ArgumentParser()
parser.add_argument(...)

def main(args=None):
 parsed_args = parser.parse_args(args)

This makes it easy to just test main() function by calling it with different arguments:

@pytest.mark.parametrize("option", ("-h", "--help"))
def test_help(capsys, option):
 try:
 main([option])
 ...

Patch `sys.argv`

It’s also possible to patch the sys.argv with the mock arguments for testing:

Notes on "Building Systems with the ChatGPT API"

Sun, 11 Jun 2023 00:00:00 +0000

Course Link: Building Systems with the ChatGPT API - DeepLearning.AI

Introduction

Process of building an application

supervised learning: usually takes long time
- get labeled data
- train model on data
- deploy & call model
prompt-based AI: takes short time
- specify prompt and call model

Language Models

How is works:

A language model is built by using supervised learning to repeatedly predict the next word.

Two types of LLMs

Full content RSS in Hugo

Sun, 04 Jun 2023 00:00:00 +0000

By default, Hugo only displays a summary of the post.

<description>{{ .Summary | html }}</description>

To enable a full text RSS feed for a section of my site, I added a layouts/section/section.rss.xml template file according to Hugo’s layout lookup order . Fill it with the default RSS layout template .

Modified the description part and add RSS content :

<description>{{ with .Description | html }}{{ . }}{{ else }}{{ .Summary | html }}{{ end -}}</description>
<content:encoded>{{ (printf "<![CDATA[%s]]>" .Content) | safeHTML }}</content:encoded>

Verify the change by going to https://localhost:1313/<section>/index.xml.

Build LLVM with CMake

Sat, 03 Jun 2023 00:00:00 +0000

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.

Using the pre-built distribution

In most cases, we can just directly download and use the pre-compiled version of LLVM and Clang from llvm-project/releases . For example, clang+llvm-16.0.0-x86_64-linux-gnu-ubuntu-18.04.tar.xz is a pre-built distribution for Ubuntu 18.04 and for the x86-64 platforms.

Build LLVM from source

In cases where there’s no pre-built version that suffice our needs, we can download and build LLVM from scratch. For example, openai/triton Python library requires a LLVM compiler to build the wheel.

Customize GitHub Codespace using Dev Container

Thu, 01 Jun 2023 00:00:00 +0000

A Development Container (or Dev Container for short) allows you to use a container as a full-featured development environment.

GitHub Codespace is a development environment that’s hosted in the cloud, which allows you to use VS Code directly from the browser to edit, commit and run projects. Whenever you work in a codespace, you are using a dev container on a virtual machine.

The dev container configuration is located under .devcontainer/devcontainer.json. There are a number of pre-built dev container images published under mcr.microsoft.com/devcontainers on devcontainers/images .

Build multi-arch Docker images in GitHub Actions

Wed, 10 May 2023 00:00:00 +0000

Publishing images to GitHub Packages

Let’s say we have a repository that includes a Dockerfile. We can utilize a GitHub workflow to:

Check out the repository
Log in to the GitHub Container Registry (ghcr.io)
Extract metadata
Build and push the Docker image to our specified registry

name: Create and publish a Docker image

on:
 push:
 tags:
 - 'v*'

env:
 REGISTRY: ghcr.io
 IMAGE_NAME: ${{ github.repository }}

jobs:
 build-and-push-image:
 runs-on: ubuntu-latest
 permissions:
 contents: read
 packages: write

 steps:
 - name: Checkout repository
 uses: actions/checkout@v3

 - name: Log in to the Container registry
 uses: docker/login-action@65b78e6e13532edd9afa3aa52ac7964289d1a9c1
 with:
 registry: ${{ env.REGISTRY }}
 username: ${{ github.actor }}
 password: ${{ secrets.GITHUB_TOKEN }}

 - name: Extract metadata (tags, labels) for Docker
 id: meta
 uses: docker/metadata-action@9ec57ed1fcdbf14dcef7dfbe97b2010124a938b7
 with:
 images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}

 - name: Build and push Docker image
 uses: docker/build-push-action@f2a1d5e99d037542a71f64918e516c093c6f3fc4
 with:
 context: .
 push: true
 tags: ${{ steps.meta.outputs.tags }}
 labels: ${{ steps.meta.outputs.labels }}

Build and publish multi-arch images

Depends on the type of the host machines, images can be built for multiple platforms, e.g. linux/amd64, linux/arm64/v8, etc.

Caddy log filter

Tue, 07 Feb 2023 00:00:00 +0000

Caddy is a powerful, enterprise-ready, open source web server with automatic HTTPS written in Go.

It can be used as a reverse proxy server, for example:

:10001 {
 respond /healthz "OK" 200
 reverse_proxy /* http://localhost:8000 {
 header_up x-account "user"
 header_up x-secret-key "secret"
 }
 log
}

The above Caddyfile will launch a http server listening on port 10001 and will redirect requests to localhost:8000 and add two headers which contain sensitive information. When the debug mode is turned on, caddy’s internal logger will log down the requests from the reverse proxy to the localhost server, which will expose the sensitive info.

Shortcut for creating daily note

Sat, 28 Jan 2023 00:00:00 +0000

I like to jog down things on my Apple Note as it’s easier and fast to access whenever my iPhone is around. With iOS built-in Shortcuts app, creating a daily note entry can be just one click.

Here’s the shortcut I made:

Format current date to be yyyy-MM-dd
Find in all notes where the name contains such date string
If exists, open the note
If not, create a note with the formatted date as title

Credit:

Decode Flask session cookie

Wed, 04 Jan 2023 00:00:00 +0000

Snippet to decode Flask session cookie.

To find the cookie in Chrome, open Inspect -> Application -> Cookies, then find the session cookie.

Here’s a short Python snippet that decodes the session cookie using zlib and base64.urlsafe_b64decode.

import zlib
import base64

def decode(cookie):
 """Decode a Flask cookie."""
 try:
 compressed = False
 payload = cookie

 if payload.startswith('.'):
 compressed = True
 payload = payload[1:]

 data = payload.split(".")[0]

 data = base64.urlsafe_b64decode(data)
 if compressed:
 data = zlib.decompress(data)

 return data.decode("utf-8")
 except Exception as e:
 return "[Decoding error: are you sure this was a Flask session cookie? {}]".format(e)

Reference:

Shortcut to copy url as markdown

Tue, 03 Jan 2023 00:00:00 +0000

This shortcut can copy a link from Safari share sheet and copy as Markdown format [title](url] to the clipboard.

Markdown reference-style link

Fri, 30 Dec 2022 00:00:00 +0000

Reference-style links are a special kind of link that make URLs easier to display and read in Markdown.

Instead of inline link: [text](link), the reference-style:

[text][label]

[label]: https://google.com "Google"

First part of the link is formatted with two sets of brackets.

The second part of a reference-style link is formatted with the following attributes:

The label, in brackets, followed immediately by a colon and at least one space (e.g., [label]: ).
The URL for the link, which you can optionally enclose in angle brackets.
The optional title for the link, which you can enclose in double quotes, single quotes, or parentheses.

See it in action:

Fn Option Delete

Fri, 16 Dec 2022 00:00:00 +0000

As long time Mac user, I wasn’t aware that “forward delete” and “backward delete” exist 😂

Backward delete option + delete

hello world▐

hello▐

Forward delete fn + option + delete

hello ▐world

hello▐

Alternatively, I was using vim editing mode for quick cursor navigation and editing.

Lightroom backup workflow

Tue, 13 Dec 2022 00:00:00 +0000

Midway through 2022, I began using Lightroom Classic to manage and edit camera-captured images. Now that the end of the year is approaching, I’ve decided to organize and back up the photos from this year on my Lacie portable hard drive so I can free up space on my MacBook Pro.

Setup

All of my photos are stored in three distinct locations:

MacBook Pro (500GB): ad-hoc editing
Sandisk Extreme Portable SSD (1TB): bulk editing
LaCie Portable Hard Drive (4TB): backup

I usually bring my laptop with me when I travel. After each day I would import my photos directly into my laptop and edit them right away. Sometimes, I import the photos to the SSD drive. LaCie hard drive is mostly used for making backups.

Merge changes from GitHub template repository

Tue, 06 Dec 2022 00:00:00 +0000

It’s very convenient to use GitHub template repository feature to bootstrap a repository. I thought it would also have functionality like Syncing a fork , but unfortunately it doesn’t.

So to sync changes from upstream template repository, we need to use git command line:

Add remote
```
git remote add template <repo>
```
Fetch template changes
```
git fetch --all
```
Use --allow-unrelated-histories to merge, we may also need to manually resolve all the merge conflicts.

Manage GitHub project in Linear

Sun, 04 Dec 2022 00:00:00 +0000

While I was working on my simple personal project: issues-blog , I experimented with Linear to manage this small project.

Linear surprised me by being a powerful and elegant tool for managing issues and projects. It has nice and modern UI/UX, and more responsive than JIRA.

Connect Linear with GitHub

Linear comes with integration with GitHub of course: GitHub - Linear Guide . To enable it, go to Settings, and under Integrations -> GitHub we can connect Linear with GitHub pull requests.

College

Mon, 01 Jan 0001 00:00:00 +0000

Brief summary of my experiences and projects during college 👨‍🎓

Internships

Uber ATG, Machine Learning Platform, 2020 Summer

Pipeline and parallelize CI/CD for inference benchmarking and evaluation.
Microsoft Research Asia, Software Analytics Group, 2019 Spring

Research on analyzing and assessing data visualizations with deep learning models.
Xiaomi Technology, AI & Cloud Platform, 2018 Fall

Model inference acceleration for production using NVIDIA TensorRT.
University of British Columbia, 2018 Summer

Uses

Mon, 01 Jan 0001 00:00:00 +0000

👨🏻‍💻 Here’s a list of tech gears, hardware, and apps I use on a regular basis.

Last updated: May 2026

Hardware

Computer

14" MacBook Pro (M1 Pro 32GB + 512GB, 2021)
IQUNIX MQ80
HHKB Professional HYBRID Type-S
Magic Trackpad
BenQ RD280UG Monitor

Audio

Apple AirPods Pro 2
Bose Noise Cancelling Headphones 700 (Soapstone)

Gaming / Accessories

Nintendo Switch 2
PlayStation 5
Kindle Paperwhite 4 (2018)
WHOOP 5.0

Software

Coding

Codex
Claude Code
GitHub

Services

1Password
Spotify
Raindrop.io
ChatGPT
Claude

Tools

Obsidian
Ghostty
Things
Anki

Xin Fu

Managing windows effectively with AeroSpace

Television: a modern fuzzy finder for the terminal

Use custom LLM providers in Claude Code

Set Google search AI mode as browser default

Run Llama 2 locally on MacBook

Download and compile llama.cpp

Create interactive utility app with Streamlit

Postboy - a simple Postman-like app

Handling JWT in Python

JWT Structure

Header

Question answering over documents with LLM

Overview

Deploy Ghost instance to Fly.io

Get Started

Custom middleware for FastAPI application

Implement custom middleware

Notes on "Langchain for LLM Application Development"

Introduction

Models, Prompts and Output Parsers

OpenAI API

LangChain

Prompt template

Set up Hugo with Tailwind CSS in 2023

How does it work?

Write pytest tests for argparse

Call main with optional arguments

Patch sys.argv

Notes on "Building Systems with the ChatGPT API"

Introduction

Language Models

Full content RSS in Hugo

Build LLVM with CMake

Using the pre-built distribution

Build LLVM from source

Customize GitHub Codespace using Dev Container

Build multi-arch Docker images in GitHub Actions

Publishing images to GitHub Packages

Build and publish multi-arch images

Caddy log filter

Shortcut for creating daily note

Decode Flask session cookie

Shortcut to copy url as markdown

Markdown reference-style link

Fn Option Delete

Lightroom backup workflow

Setup

Merge changes from GitHub template repository

Manage GitHub project in Linear

Connect Linear with GitHub

College

Internships

Uses

Hardware

Computer

Audio

Gaming / Accessories

Software

Coding

Services

Tools

Call `main` with optional arguments

Patch `sys.argv`