Projects

### Overview: Gamba is a PyTorch implementation of the GAMBA algorithm, designed for single-view 3D reconstruction. This project aims to marry Gaussian splattering with Mamba for efficient and accurate 3D reconstruction. ### Key features: - **GAMBA Algorithm Implementation**: Implements the GAMBA algorithm for single-view 3D reconstruction. - **PyTorch Framework**: Utilizes PyTorch for seamless integration and performance optimization. - **Efficient Reconstruction**: Utilizes Gaussian splattering and Mamba techniques for fast and accurate reconstruction. ### Tech stack: Python ### Why it matters: Gamba provides a practical and efficient solution for single-view 3D reconstruction, leveraging advanced techniques to achieve high-quality results with minimal computational resources.

Overview: This repository contains a pre-trained Electra small model trained on a large Vietnamese corpus (~50GB of text). Inspired by generative adversarial networks (GANs), ELECTRA trains the model to distinguish between "real" and "fake" input data by replacing some input tokens with incorrect but plausible fakes, rather than corrupting the input as in BERT. Key features: - Pre-trained Electra small model - Vietnamese corpus training - Tokenization using coccoc-tokenizer Tech stack: TensorFlow 2.1.0 Why it matters: This pre-trained model can be used for various NLP tasks, such as text classification, summarization, and translation, by leveraging the Vietnamese corpus for better performance and context-aware representation.

### Overview: OpenVenice is an open-source frontend for the Venice AI API, providing a customizable and transparent interface for AI model usage. ### Key features: - **Customizable Interface**: Modify the UI to suit your needs, including adding custom tools and parameters. - **API Key Management**: Store your API key securely in your browser without needing separate accounts. - **No Backend**: Host OpenVenice on your own server, and your API key is accessible directly via the browser. - **Transparent and Minimal**: Every API call is visible in the source code, and no telemetry or analytics are used. ### Tech stack: TypeScript ### Why it matters: Your interface, your rules. Customize everything, share your API key with family, no server or backend required, and transparent and minimal features for a clean and efficient experience.

### Overview: CueRecord is a native macOS screen recorder that includes a teleprompter, camera overlay, system audio capture, and AI-assisted script preparation. It offers high-resolution 4K export and AI breath cuts for polished product demos, tutorials, courses, and presentations. ### Key Features: - **Screen Recording**: Native Swift macOS app for full-screen, window, and selected-area recording. - **Camera Overlay**: Optional camera overlay with draggable preview window. - **Microphone and System Audio Capture**: Integrated for capturing audio. - **AI Breath Cuts**: Adds natural teleprompter line breaks using OpenAI-compatible models. - **Built-in Teleprompter**: Adjustable line count and follow-along behavior. - **Export Presets**: Resolution presets up to 4K with visible expected output dimensions. - **Raw Recording Assets**: Organized in `raw_data/` while the final composited movie stays in the project folder. - **Developer ID Signed DMG**: Distribution-ready with minimal export progress in the project sidebar. ### Tech Stack: Swift ### Why It Matters: CueRecord is designed for creators, educators, founders, and developers who need to record polished walkthrough videos without the hassle of multiple tools. It combines high-quality screen recording, teleprompting, audio capture, and AI-assisted script preparation in one integrated solution. Repository: nolangz/CueRecord Primary Language: Swift GitHub Stars: 65 GitHub Description: Native macOS screen recorder with teleprompter, camera overlay, system audio capture, 4K export, and AI

Overview: This AI plugin allows users to generate various artifacts from their Google NotebookLM notebooks, including slide decks, audio overviews, videos, mind maps, flashcards, quizzes, infographics, reports, and data tables. Key features: - Generate slide decks - Create audio overviews - Produce video overviews - Create mind maps - Generate flashcards - Create quizzes - Generate infographics - Create reports Tech stack: TypeScript Why it matters: This plugin provides a seamless integration of the full NotebookLM experience into Claude Code, OpenClaw, and other AI agents that execute shell commands, making it easier to manage and generate artifacts from notebooks.

### Overview: This repository contains notebooks and use cases for fine-tuning state-of-the-art Vision-Language Models (VLMs) on custom datasets. All examples are designed to run on free Google Colab notebooks. ### Key features: 1. Image Understanding: VLMs can analyze and interpret complex visual scenes. 2. Text Generation: Based on visual inputs, these models can generate relevant textual descriptions or responses. 3. Multi-modal Learning: VLMs can process and correlate information from both visual and textual sources. ### Tech stack: Jupyter Notebook ### Why it matters: Vision Language Models (VLMs) are a cutting-edge type of artificial intelligence that combines computer vision and natural language processing. These models can understand and generate text based on visual inputs, bridging the gap between visual and textual information.

### Overview: An AI-powered tool for transcribing, summarizing, and creating smart clips from video and audio content. ### Key features: - **Transcription**: Transcribe audio using WhisperX - **Smart Summarization**: Generate concise summaries tailored to various purposes: - Meeting Minutes - Podcast Summaries - Lecture Notes - Interview Highlights - General Content Summaries - **Intelligent Clip Creation**: Automatically create clips of key moments and topics discussed in the video - **Multi-format Support**: Process various video and audio file formats - **Cloud Integration**: Utilizes AWS S3 for efficient file handling and processing ### Tech stack: Python ### Why it matters: This tool streamlines the process of summarizing and summarizing video content, making it easier for developers to quickly extract key information and insights from their work.

Overview: Luna is a text-to-image generator using Stable Diffusion, capable of generating detailed images based on text descriptions. It can also be applied to tasks like inpainting and image-to-image translation. Key features: - Text-to-image generation - Supports inpainting and image-to-image translation - Runs on both CPU and GPU - Available online on Google Colab for quick access Tech stack: Python Why it matters: Luna offers a powerful tool for generating detailed images based on textual prompts, making it valuable for various creative and technical applications. Repository: slowy07/luna GitHub stars: 65 GitHub description: text to image generation with stable diffusion

### Overview: This repository serves as a collection of resources and materials for beginners interested in learning about Large Language Models (LLMs). It includes a mix of theoretical and practical resources, covering topics such as NLP fundamentals, Transformers, and large language models. The goal is to provide a structured roadmap for learning LLMs in a proper manner. ### Key Features: - 📖 NLP Fundamentals: Resources on NLP, including courses from Stanford and Coursera. - 🤖 Transformers: Detailed course material on Transformers from HuggingFace. - 🦜 Large Language Models: Comprehensive course on LLMs, including courses from Activeloop, Maxime Labonne, and others. ### Tech Stack: - Python, Jupyter Notebooks, and various programming languages used for coding and experimentation. ### Why It Matters: A clear roadmap and structured learning resources will help beginners effectively understand and apply LLMs in practical scenarios. It serves as a foundational guide for anyone looking to delve into the world of LLMs. ### Repository: snehilsanyal/self-learn-llms - GitHub Stars: 65 - Description: A collection of resources, notebooks, blogs, and tutorials for beginners learning about Large Language Models. The initiative aims to provide a comprehensive and visual learning experience.

Overview: This repository contains the source code for the iOS client library for Speechly API, allowing developers to easily integrate voice-enabled applications with Speechly's SLU API. Key features: - Speechly iOS Client Library - Swift package dependency - Xcode package dependency - Delegate protocol for receiving recognition results - Start and stop recognition context methods Tech stack: Swift Why it matters: This library simplifies the integration of voice-enabled applications with Speechly's SLU API, making it easier for developers to build voice-driven user interfaces without the need for complex configuration or manual setup.

Overview: This module provides a robust way to render text and HTML content into images using headless Chrome, making it easy to achieve high-quality output for various applications. Key features: - Built-in fontface observer - Loads Google fonts easily - Option to apply word-wrap - Renders HTML content with custom styling - Supports multiple fonts - Thoroughly tested - Includes a CLI version Tech stack: JavaScript Why it matters: This module simplifies the process of rendering text and HTML into images, providing a more modern and efficient alternative to traditional methods like ImageMagick and Pango.

Overview: OFTSR is an implementation of a paper that introduces a novel approach for image super-resolution, specifically focusing on FFHQ and DIV2K datasets. The method allows for tuning the fidelity and realism of the restored images. Key features: - FFHQ Noiseless Restoration - FFHQ Noisy Restoration - DIV2K Noiseless Restoration - ImageNet Noiseless Restoration Tech stack: Python Why it matters: This method offers a one-step flow for image super-resolution, providing flexibility in balancing fidelity and realism, making it a valuable tool for enhancing images in various applications.

Overview: This sample Unity project demonstrates speech recognition using the new Microsoft Speech Service (Preview) via WebSockets. Key features: - Speech recognition using the new Microsoft Speech Service (Preview) - Real-time speech recognition results posted in the Unity UI and Debug Console - Compatible with both the new Speech Service and classic Bing Speech API - Supports two methods for speech recognition: file upload and microphone recording Tech stack: C# Why it matters: This project showcases the integration of Microsoft's new speech recognition service into Unity, providing a practical demonstration of how to use it in a game environment.

Overview: EDG4LLM is a Python library that generates fine-tuning datasets for large language models, including questions, answers, and dialogues. Key features: - Generate question data - Generate answer data - Generate dialogue data Tech stack: Python Why it matters: EDG4LLM simplifies the process of creating high-quality training datasets for large language models, making it easier for developers to fine-tune models effectively.

### Overview: VRCTextboxSTT is a SpeechToText application that uses OpenAI's Whisper via faster-whisper to transcribe audio and send the information to VRChat's textbox system or KillFrenzyAvatarText over OSC. It supports various methods such as OBS via BrowserSource and a SteamVR overlay, and it is designed to be completely free, open-source, and independent from cloud-based transcription services. ### Key Features: - Transcription to VRChat Ingame Textbox or KillFrenzyAvatarText - Support for OBS over BrowserSource - SteamVR Overlay for viewing transcription without looking at the textbox - Fast and Efficient Transcription using ctranslate2 - Customizable Transcription Modes and Timings - Automatic Launch with SteamVR - Customizable Audio Feedback - Multi-Language Support - Word Replacements and Emote Replacements - Simple API for Transcription and Text-to-Text ### Tech Stack: Python ### Why it Matters: This program is designed to be completely free, open-source, and independent from cloud-based transcription services such as Microsoft Azure. It utilizes transcription algorithms that run on your own hardware, ensuring privacy, enhancing latency, and ensuring reliability.

Overview: hfgo is an unofficial Go SDK designed for interacting with the Hugging Face Inference API, providing type-safe client methods for various tasks and a pluggable provider layer for alternate wire formats. Key features: - Type-safe client methods for every supported inference task - Pluggable provider layer for alternate wire formats Tech stack: Go Why it matters: hfgo simplifies interaction with the Hugging Face Inference API, making it easier and safer to integrate with the API using Go.

### Overview: S.T.A.R.K. - Speech and Text Advanced Recognition Kit is a modern, async Python framework designed for building voice assistants and natural language interfaces. It provides a robust, on-device solution that ensures your data stays secure and your application remains functional even when the internet connection is lost. The framework includes deterministic, fast, and explainable pattern, phonetic, and fuzzy matching, making it perfect for applications requiring high accuracy and reliability. ### Key Features: - **100% On-Device**: Runs fully offline, ensuring your data stays yours and your application remains operational even without an internet connection. - **4 Required Dependencies**: `pydantic`, `asyncer`, `anyio`, `numpy`. Everything else is opt-in, allowing for flexibility in adding speech-to-text and text-to-speech backends. - **No AI Required**: Pattern, phonetic, and fuzzy matching are deterministic, fast, and explainable, making it suitable for applications that require high accuracy and reliability. - **Phonetic and Fuzzy Matching**: Handles misspellings, accents, and cross-language name lookup out of the box. - **Multilingual by Design**: Supports commands that mix languages mid-sentence, making it ideal for applications that need to support multiple languages. - **Nested Contexts**: Multi-level menus, follow-ups, and stateful conversations, providing a more engaging user experience. - **Background Commands & Multiple Responses**: Allows for fire-and-forget tasks, keeping the assistant active and ready to respond. - **Assistant Modes**: Active, waiting, inactive, sleeping

Overview: MorphingDB is a PostgreSQL extension designed to support deep learning model inference within the database and vector storage, enabling efficient inference and vector storage for models. Key features: - Import and run LibTorch models within PostgreSQL. - Store vector data with dimension information for LibTorch to work directly with. - Support both model-centric and task-centric inference. - Quick start guide with Docker, zip download, and SQL examples. Tech stack: Python, PostgreSQL, LibTorch, Docker. Why it matters: This extension simplifies the integration of deep learning models into traditional database environments, enhancing data processing and storage capabilities. Repository: MorphingDB/MorphingDB GitHub stars: 64 GitHub description: PostgreSQL extension for supporting deep learning model inference within the database and vector storage.

Overview: LORIS is an implementation of "Long-Term Rhythmic Video Soundtracker" presented at ICML 2023. It is a novel framework for synthesizing long-term conditional waveforms in sync with visual cues, extending its applicability from dances to multiple sports scenarios. Key features: - Latent conditional diffusion probabilistic model for waveform synthesis. - Context-aware conditioning encoders for temporal information consideration. - Extends model's applicability from dances to floor exercise and figure skating. - Established benchmark for rhythmic video soundtracks with pre-processed dataset, improved evaluation metrics, and robust generative baselines. Tech stack: Python Why it matters: LORIS provides a comprehensive solution for generating rhythmic video soundtracks, offering significant improvements in both research and practical applications. Repository: OpenGVLab/LORIS Primary language: Python GitHub stars: 64 GitHub description: [ICML2023] Long-Term Rhythmic Video Soundtracker [Arxiv](https://arxiv.org/abs/2305.01319) | [Project Page](https://justinyuu.github.io/LORIS/)

### Overview: This is a production-serving kit for GLM-5.3-Flash (320B MoE, 18B active) running on two NVIDIA DGX Spark machines. It includes 1,000,000-token context with DFlash2 speculative decoding and tensor-parallel 2. ### Key features: - 1,000,000-token context - DFlash2 speculative decoding - Tensor-parallel 2 - 3,584-token pages in KDA state - 100% multi-session prefix caching - 97%+ multi-session prefix caching ### Tech stack: Python ### Why it matters: This kit optimizes for agentic traffic, providing 97%+ multi-session prefix caching, reducing prefix-cache misses and prefill latency, and improving performance by 98%+ under real agentic traffic.