<RAG> 2. Collecting RAG Data Using LLM

Friday, May 16, 2025

Scraping data will give us large volume of data in small time, but we need a way more structured type of data for the model to process if we’re going to use this data in the RAG dataset. Summarizing this data by hand will take a very long time, so we will use the LLM to summarize our scraped data for us. Coming up with a good data structure Funnily, one of the best usage of LLM is to as how to better use an LLM in a project.

<RAG> 3. Setting Up Baseline LLM Prompt For Service

Friday, May 16, 2025

While Gemini summarizes my dataset, I’ll be creating an LLM server skeleton to use in the project. Backend Model Structure We’re going to provide an LLM Adapter so that we can easily transfer from local smaller LLM and larger LLM hosted by bigger companies LLM Request Adapter """Provides API for requesting Online and Offline LLM Character Evaluation """ from abc import abstractmethod class LlmCharacterEvalAdapter: @abstractmethod def get_character_evaluation_from_llm(self, character_eval_request) -> CharacterEvalResponse: """Request Character Evaluation to the target LLM with the given request format.

<Deep Learning Intermediate> Optuna for Effective Hyperparameter Tuning

Sunday, May 4, 2025

Hyperparameter tuning is an important part of training ML models. Machine learning engineers used to tune hyperparameters by hand or use Grid Search. These days there are more powerful tools that can help tune hyperparameters in less time. We’re going to look into a easy-to-use tool called Optuna Problems With Grid Search Takes too long: Search space gets huge as more hyperparameters get inside grid. If there is 86,000 grids already, adding one hyperparameter with just binary options doubles the number of grids to 172,000 Doesn’t utilize previous search result: As we do the search, we can get a sense of which range of hyperparameters are more effective to experiment Only supports discrete grids ex) After doing 10 experiments with learning rate between 0.

<RAG> 1. Tuning LLM Using Langchain + LM Studio

Friday, May 2, 2025

I’m working on a story evaluation AI Agent. I’m going to tune the model using Langchain with a local LLM using LM Studio 1. Setting up LM Studio Tool for downloading and running local LLM model that fits your local environment. Download: https://lmstudio.ai/ Install and we’ll be asked to download Qwen B7 for onboarding. After downloading, we can run the model Run the model and we’ll se a chatbot environment pop up.

<C++ Server Programming> Troubleshooting - C++ Server and Unreal Engine

Monday, Apr 28, 2025

1. nested threads in C++ Situation I was writing a unit test for my GlobalThreadManager Singleton Class: TEST_F(ThreadManagerTest, ThreadIdGeneration) { std::vector<std::thread> jobs; std::atomic<int> completion_counter{ 0 }; // Given: 10 threads are generated asynchronously by 10 different threads using GlobalThreadManager, total 100 threads for (int i = 0; i < 10; i++) { jobs.push_back(std::thread([this, &completion_counter]() { for (int j = 0; j < 10; j++) { GlobalThreadManager->Launch([this, &completion_counter]() { std::lock_guard<std::mutex> lock(thread_id_mutex); thread_ids.

<Deep Learning Basics> ML Basics - GPU Types and Performance

Monday, Apr 28, 2025

To be a good machine learning engineer you have to know what types of GPUs there are and their relative performances. 1. Data Center GPUs (AI/HPC) The best GPU types and optimized for AI/HPC. They are usually named in AX00, where A is the Capital letter of their architecture, and X is a number that represents improvement in the current architecture ex) _ H100: First version of Hopper GPU _ H200: Second version of Hopper GPU

<C++> Installing whisper.cpp

Monday, Apr 14, 2025

I’m going to make a game that interacts with NPCs with voice. To do that I need to install a voice recognition AI model which translates user’s voice into natural language. OpenAI has a model called whisper which does exactly this, and it has a community C++ extension(https://github.com/ggml-org/whisper.cpp ) I’m setup this whisper.cpp in my game file so that we can run it locally. 1. Clone Repository and Download Model Weights $ git clone https://github.

<Server> Understanding Game Servers(C++)

Monday, Apr 14, 2025

Web Server vs Game Server Web Server: small # of requests per user don’t need real-time interaction Server doesn’t approach client first Stateless - forgets user after request is done Usually uses web frameworks ex) drivethrough Game Server: a lot of request/response real-time interaction stateful cannot use web framework (different needs for each game) ex) restaurant - constant service to the user C++ Thread #include <thread> #include <thread> int main() { std::thread t; // CPU core number int t.

Python OOP Features

Saturday, Apr 12, 2025

Studying on OOP Features of Python Programming Language 1. Class Method for Polymorphism class Animal: def __init__(self, name): self.name = name def say(self): pass class Cat: def __init__(self, name): super().__init__(name) def say(self): print(f"{self.name}, a Cat") class Dog: def __init__(self, name): super().__init__(name) def say(self): print(f"{self.name}, a Dog") class AnimalContainer: def __init__(self): self.animals = [] def create_animals(self, names): for name in names: animal = Cat(name) animal.say() self.animals.append(animal) With the current implementation, if you want to implement create_animals function for Dog, you have to create a new class or a new function for Dogs.

<Analytics> Starting Spark Session in Azure Synapse Analytics

Sunday, Mar 30, 2025

I was doing a project helping an environmental organization find ways to better store and utilize their geospatial data(mostly .shp files and point clouds). As a solution we recommended moving their data from Dropbox to Azure (Blob Storage) for more flexible storage management and options to analyze data. To give a proof-of-concept about the benefits of using cloud storage, I made a geospatial analytics prototype using Azure Synapse Analytics. 1. Creating Synapse Analytics Workspace To use Synapse Analytics, we have to create a workspace to use our data.