moomou

Software for One

Tue, 28 Apr 2026 00:00:00 +0000

Introduction #

In the age of coding agents, building software has never been cheaper or easier. I can go from an idea to shipping working software in an afternoon. I wanted to document a couple of these in this post and argue that, in the era of AI coding, more software will be written and increasingly cater to a smaller audience.

WinGroup #

WinGroup is a Mission Control–style window overlay and a per-app hotkey launcher. It replaces Apptivate, the last Rosetta-translated app on my Mac, and adds two features I could not find in any single shipping app: AI auto-grouping of open windows by project or task, and a way to move a whole group of windows to another desktop space in one action.

Building chef-py: Replacing Chef with Python

Thu, 19 Mar 2026 00:00:00 +0000

Introduction #

I’ve been using Chef for server configuration for a while now. It works, but there’s always been friction. I don’t write Ruby anywhere else, so every time I need to update server configs, I’m context-switching into a language I’m barely literate in. Chef uses Ruby with its own DSL conventions, and while that’s arguably better than learning a completely custom language like Puppet, I still find myself fighting against patterns I don’t understand and don’t use anywhere else.

Building Flow Matching Models from Scratch

Tue, 10 Mar 2026 00:00:00 +0000

Introduction #

Continuing the path of learning about deep learning, I have been working on building a diffusion/flow matching models from scratch on and off for a while. My initial goal was to build a flow-matching or diffusion model for audio with mel-spectrograms inputs. However, after struggling with eyeballing generated melspectrogram to determine if it’s good or bad and the added complexity of dealing with melspectrogram to audio conversion, I decided to pivot to the image domain first to make sure I learn the diffusion modeling aspect properly before tackling audio. Specifically, I chose to go with with the classic celebrity datasets, allowing me to focus on the core mechanics of generation.

Learning about CT with Gemini

Wed, 07 Jan 2026 00:00:00 +0000

Overview #

Over the winter break, I was reviewing Fourier Transform and its various applications. Fundamentally, Fourier Transform is about a shift in perspective - viewing the same data with a different lens makes some operations easier or faster. This piqued my curiosity about what other fundamental data transformations that I should be aware of. Like everyone in 2025, I jumped to my chatbot of choice, Gemini, to learn more.

Navigating within a Mono Repo

Mon, 29 Dec 2025 00:00:00 +0000

Groot and Friends #

Navigating within a mono repo can be tedious.

I frequently find myself doing cd ../.. to navigate to sibling folders or jumping to the root of the git directory to navigate and inspect other part of the code base.

To navigate to the root of the Git repository, I use an alias groot defined as

    alias groot='cd $(git rev-parse --show-toplevel)'

Excellent tools such as fd and fzf help a lot but I want to be able to jump to different part of the repo with a few keystrokes. For example, in a mono repo, I want to be quickly jump from go/src/mou.dev/cmd to schema/proto.

Listening with LLM

Sun, 31 Dec 2023 00:00:00 +0000

Overview #

This is the first part of many posts I am writing to consolidate learnings on how to finetune Large Language Models (LLMs) to process audio, with the eventual goal of being able to build and host a LLM able to describe human voices.

I am motivated to gain hands-on experience tinkering LLMs so, as much as practical, I tried to recreate utilities and functions with pytorch from scratch rather than rely on 3rd party libraries.

Learning to Debug Gibberish

Wed, 27 Dec 2023 00:00:00 +0000

Overview #

Recently, I went down a rabbit hole of debugging opensource LLM spewing gibberish on my PC. The investigation was the one of the most difficult (but interesting) debugging experience I have encountered so far.

This blog will chronicle the problem I encountered, debugging steps I took, and finally, the resolution and workaround I adopted.

tl;dr I traced down the source of LLM spewing gibberish to faulty RAM sticks after getting different sha256 checksums on a static file on different runs. I was then able to salvage ~40% of the faulty RAM stick capacity after discovering Linux kernl parameter memtest=.

CORS with Cookie

Sat, 13 Aug 2022 00:00:00 +0000

Introduction #

CORS stands for Cross-Origin-Resource-Sharing and is the HTTP mechanism to allow servers to accept requests from other host locations other than its own.

Getting CORS up and running is not difficult. There are plenty of posts online that will recommend a starting configuration that looks something like this, in pseudo code

Header().Set("Access-Control-Allow-Origin", "*")
Header().Set("Access-Control-Allow-Methods", "POST, GET, OPTIONS, PUT, DELETE")
Header().Set("Access-Control-Allow-Headers", "Accept, Content-Type, Content-Length, Authorization")

This configuration works great and is the default I use; however, it does not work if you need to receive and set cookies in the browser. This short blog post documents what I learnt about working with cookies with CORS.

Messing with Nvidia GPU on Headless Linux

Wed, 29 Dec 2021 00:00:00 +0000

I have a PC with Ubuntu server installed and Nvidia GPUs attached. I have Googled on and off for a while trying to learn how to overclock the GPUs without success.

Finally, after gleaning bits and parts from different sources, I got power limit and overclocking working.

Here are the steps required

Virtual Monitor #

nvidia-settings need a monitor attached in order to work. For headless linux, one needs a virtual monitor. After messing with xorg conf manually for a while, I found and forked andyljones/coolgpus to automatically attach a virtual monitor for each GPU.

Improvements to "m", a personal command line tool

Sat, 22 Feb 2020 00:00:00 +0000

From m to m2 #

Having used m for many years now, I encountered two primary problems.

The first is startup performance. Even with lazy module loading (see here), calling m without any subcommands still took ~250ms on my laptop.

The second problem is code verbosity. To add a new command, one needs to create a new class in a separate file, import it in main, and finally instantiate the new class.

Training a Speaker Embedding from Scratch with Triplet Learning

Sat, 05 May 2018 23:01:25 -0700

Posted on go.mou.dev/triplet-embedding-learning

ML paper notes

Fri, 01 Sep 2017 23:01:25 -0700

2017-09 #

LEARNING FINE-GRAINED IMAGE SIMILARITY WITH DEEP RANKING #

describes efficient sampling technique based on reservoir sampling for building triplets; requires an relevance function
multi scale CNN

DEEP METRIC LEARNING USING TRIPLET NETWORK #

learns a semantic embedding; results show better discrimination vs siamese network (contrastive loss function)
MSE softmax shows improved performance rather than simple binary softmax (see paper for def)
feed a triplet of x, x1, x2 where x1 is same class as x and x2 is different

DISTILLING THE KNOWLEDGE IN A NEURAL NETWORK #

explores compression technique of ensemble model into a single model
Distillation
- softmax qi = exp(zi/T)/Sigma(j)(exp(zj/T) where z are logits and T is temperature
- T is usually 1
- increasing T creates softer probability distribution
- knowledge is tranferred via training smaller/compressed model by targeting over softer target (ie temperature T > 1) from more cumbersome model
- small model trained with higher T as well but in prediction mode uses T = 1
- tranfer training can be improved by using datasets with true label
demonstrate distillation with minist dataset - tranfer works well even when smaller model trained by omitting certain numbers
discusses using soft distribution target technique for training specialists on very large datasets
Google internal JFT data of 100M images

Questions #

teacher - student model, relation to curriculum learning?

Tools

Fri, 01 Sep 2017 23:01:25 -0700

Fzf #

Amazing command line tool to fuzzy search for files. Cannot live without this. Also integrates with vim.

https://github.com/junegunn/fzf

Jq #

Swiss army knife for working with JSON on the command line

https://github.com/stedolan/jq

Pup #

Like Jq, but for HTML

https://github.com/ericchiang/pup

Silversearcher #

Like grep, but better

https://github.com/ggreer/the_silver_searcher

Developer Experience

Tue, 22 Aug 2017 23:01:25 -0700

Commandline Productivity and Automation (aka make it easy to repeat) #

Developers tend to repeat themselves. A lot. This can be as innocuous as running a test manually after you update a test file or as insidious as deploying a newly built binary into production manually. Nobody likes to repeat themselves over and over again. And for good reasons. Not only does it take time to repeatedly figure out all the parameters in a tricky command line program, it also increases stress when you are trying to fix outages as quickly as possible.