Projects

### Overview: FeatherSpec is a lightweight, spec-driven development template designed for Claude Code, GitHub Copilot, and JetBrains Junie. It leverages Markdown and folders to manage specifications, plans, and code, ensuring that the workflow is seamless across different tools. ### Key Features: - **Spec-driven Development**: Describes what you want, generates a spec with testable acceptance criteria, then into a plan of baby steps, and finally writes code. - **Memory Bank**: Tracks work in Jira or GitHub Issues, linking specs to tickets and statuses. - **No CLI, No Install, No Dependencies**: Uses Markdown and folders for simplicity. ### Tech Stack: - **Markdown and Folders**. ### Why It Matters: FeatherSpec streamlines the development process by ensuring that the entire workflow is consistent across different tools, making it easier for teams to collaborate and maintain a clean, organized codebase.

Overview: This project introduces and demonstrates the practical applications of Retrieval Augmented Generation (RAG) using Python code in a Jupyter Notebook environment. Key features: - Basic RAG Pipeline - Sentence Window Retrieval - Auto-Merging Retrieval Tech stack: Python, Jupyter Notebook Why it matters: This project serves as an educational and fun resource for developers looking to learn and experiment with RAG, providing hands-on experiments and a dedicated Python environment for seamless code execution.

### Overview: Echo is a production-ready audio and video transcription app that can run on laptops or in the cloud. Powered by OpenAI's Whisper, it offers near-human speech recognition for faster and easier captioning. ### Key Features: - **Transcription and Captioning**: Quickly generate captions for video and audio content. - **Open Source**: Built with Python and available on PyTorch Lightning. - **Production Ready**: Toggles monitoring and production-specific features. ### Tech Stack: TypeScript ### Why It Matters: Echo leverages OpenAI's Whisper to provide high-quality speech recognition, making it an efficient and reliable tool for transcribing audio and video content.

### Overview: This repository contains the official code and datasets for the paper "Combining audio control and style transfer using latent diffusion", presented at ISMIR 2024. It includes instructions for training your own model and access to pre-trained checkpoints for audio-to-audio and MIDI-to-audio transfers. ### Key features: - **Model Training**: Includes steps for dataset preparation, autoencoder training, and diffusion model training. - **Pretrained Models**: Available for audio-to-audio transfer and MIDI-to-audio transfer. - **Real-Time Implementation**: A real-time version of the model is available in the AFTER repository. ### Tech stack: Jupyter Notebook ### Why it matters: This repository provides a comprehensive solution for combining audio control and style transfer using latent diffusion, enabling creative audio manipulation and real-time applications.

### Overview: Zero-Painter is an innovative training-free framework for layout-conditional text-to-image synthesis, achieving high fidelity in preserving textual details and adhering to mask shapes. ### Key features: - **Prompt-Adjusted Cross-Attention (PACA) and Region-Grouped Cross-Attention (ReGCA) Blocks**: Enhances precise alignment of generated objects with textual prompts and mask shapes. - **Training-Free**: Facilitates the creation of detailed and controlled imagery without the need for training. - **Text-to-Image Synthesis**: Generates images from textual prompts with high fidelity. ### Tech stack: Python ### Why it matters: Zero-Painter surpasses current state-of-the-art methods in preserving textual details and adhering to mask shapes, making it a significant advancement in text-to-image synthesis.

### Overview: This project introduces HiDream O1 Image nodes and LoRA training for ComfyUI, enabling direct image generation within the ComfyUI interface. The nodes support various precision settings and model loading options, including BF16, FP16, FP32, and FP8, with a focus on reducing VRAM usage and maintaining performance. ### Key Features: - **Direct HiDream O1 Image Generation**: Generate images directly within ComfyUI using text prompts. - **Text and Reference Image Workflow**: Combine text prompts with reference images for more detailed results. - **Dynamic Inputs**: Support for `image_1` to `image_12` inputs on the sampler node for flexibility. - **Conditional Dev Layout**: Optional conditioning via JSON bbox input for more controlled outputs. - **Model Loading Options**: Support for BF16, FP16, FP32, and FP8 E4M3FN, FP8 E5M2, and FP8 mixed-weight loading. - **FlashAttention and SageAttention Backends**: Utilize FlashAttention and SageAttention for improved performance. - **Progress Previews**: Real-time progress previews through ComfyUI's sampler progress bar for better user experience. - **Dev/Dev-2604 Patch-Grid Smoothing Node**: Smooth out visible tile seams for better visual quality. - **AI Toolkit-Aligned LoRA Training**: Train LoRA models aligned with the AI Toolkit for better compatibility. - **ComfyUI Model Management, Unload, and Aimdo/VBAR Support**: Ensure smooth model management and

### Overview: Scottcjn/llama-cpp-tigerleopard is a project that provides a working port of llama.cpp for Mac OS X Tiger and Leopard running on PowerPC G4 and G5 processors. This is the first public release of this port and marks a significant milestone in bringing LLM inference capabilities to classic Mac hardware. ### Key Features: - **20-Year-Old Hardware**: Running modern LLM inference on vintage Macs. - **AltiVec/Velocity Engine SIMD Acceleration**: Utilizing SIMD for improved performance. - **No GPU Required**: Pure CPU inference, making it ideal for classic Mac hardware. - **Vintage Computing Meets AI**: Your 2005 Power Mac can now run LLM inference! ### Tech Stack: C++ ### Why It Matters: - **20-Year-Old Hardware**: Running modern LLM inference on vintage Macs. - **AltiVec/Velocity Engine SIMD Acceleration**: Utilizing SIMD for improved performance. - **No GPU Required**: Pure CPU inference, making it ideal for classic Mac hardware. - **Vintage Computing Meets AI**: Your 2005 Power Mac can now run LLM inference! ### GitHub Stars: 70 ### Repository: Scottcjn/llama-cpp-tigerleopard ### Primary Language: C++ ### Performance: - **TinyLlama 1.1B Q4**: G5 Dual 2.0GHz - ~3-5 t/s - **TinyLlama 1.1B Q4

### Overview: UnTrack is a unified multi-sensor tracker that leverages a single parameter set for any modality, enabling effective unification and accommodating missing modality within a single transformer-based architecture. This approach simplifies the tracking process and avoids the need for modality-specific fine-tuning. ### Key Features: - Unified Multi-Sensor Tracker - Single Parameter Set for Any Modality - Seamless Binding of All Modalities - Effective Unification and Accommodation of Missing Modality - Single Transformer-Based Architecture ### Tech Stack: Python ### Why It Matters: UnTrack simplifies the tracking process by unifying multiple sensor modalities into a single, unified model, making it more efficient and easier to implement compared to specialized methods for each modality. This approach reduces the need for extensive fine-tuning and improves overall performance across various applications.

Overview: This polyfill provides a modern and efficient way to use the Web Speech API with Cognitive Services Speech Services, enabling speech-to-text and text-to-speech functionalities across modern browsers. Key features: - Speech recognition - Speech synthesis Tech stack: JavaScript Why it matters: This polyfill simplifies the integration of speech technologies into modern web applications, making it easier to implement intelligent personal assistants and assistive technologies.

### Overview: This project fine-tunes 6-Billion GPT-J and other models using LoRA and 8-bit compression, providing a detailed example for fine-tuning models with limited memory. ### Key Features: - Fine-tuning 6-Billion GPT-J and other models - Utilizes LoRA (Low-Rank Adaptation) and 8-bit compression techniques - Detailed explanation found in the model card - Based on a Colab notebook - Trains models on philosophical, trippy, and mystical content - Trains models on French news ### Tech Stack: Jupyter Notebook ### Why it Matters: This project offers a practical example for fine-tuning large models with limited computational resources, making it useful for developers and researchers looking to optimize model training on constrained hardware.

Overview: This project provides a Python implementation for pre-processing techniques used in End-to-End speech recognition, including Fisher, AMI, WSJ0, WSJ1, and Chime3 datasets. Key features: - Fisher - AMI - WSJ0 - WSJ1 - Chime3 Tech stack: Python Why it matters: This project is essential for researchers and developers working on speech recognition systems, offering a robust pre-processing toolkit that can be easily integrated into various projects. Repository: hirofumi0810/asr_preprocessing Primary language: Python GitHub stars: 70

### Overview: FastWER is a Python package designed for fast calculation of word/character error rate (WER/CER) in natural language processing tasks. It utilizes a C++ implementation for efficiency and supports both sentence-level and corpus-level WER/CER scores. ### Key Features: - **Sentence-Level and Corpus-Level Scores**: Measures the error rate at the sentence and corpus levels. - **Unicode Code Point-Based Error Rates**: Accurately calculates character error rates based on Unicode code points. - **Efficient Performance**: Utilizes a C++ implementation for fast computation. ### Tech Stack: Python ### Why It Matters: FastWER is essential for developers working on natural language processing tasks where precise error rate calculations are required. Its performance and accuracy make it a valuable tool for improving text quality and ensuring data reliability in applications like machine translation, text summarization, and more. ### Repository: kahne/fastwer ### GitHub Stars: 70 ### Primary Language: Python ### GitHub Description: A PyPI package for fast word/character error rate (WER/CER) calculation

Overview: This proof of concept app showcases the use of KeenASR SDK in Objective-C, providing a glimpse into the integration of automatic speech recognition technology. Key features: - Educational Reading Demo - Educational Words Demo - Command and Control Demo Tech stack: Objective-C Why it matters: This app demonstrates the practical application of KeenASR SDK in real-world scenarios, highlighting its utility for educational and command control applications. Repository: keenresearch/keenasr-ios-poc Primary language: Objective-C GitHub stars: 70 GitHub description: Proof of concept app that demonstrates use of KeenASR SDK in ObjC. We are hiring: https://keenresearch.com/careers.html README excerpt: This proof-of-concept app shows how to run KeenASR automatic speech recognition framework. For detailed information, consult the SDK reference documentation. If starting from scratch, check our Quick Start document. Three demos are provided: Educational Reading, Educational Words, and Command and Control.

Overview: The AGIQA-3K database is a comprehensive fine-grained subjective quality assessment database for AI-generated images, providing detailed subjective scores for both perception and alignment levels. Key features: - Contains 2 types of files: AI-Generated Images (AGIQA-3K.zip) and data.csv (Prompt, perception, alignment). - Provides normalized MOS and STD scores for both perception and alignment levels. - Offers a database for the evaluation of AI-generated image quality. Tech stack: Python, OpenCV, Pandas, Matplotlib. Why it matters: The fine-grained subjective scores in AGIQA-3K will inspire subsequent AGI quality models to better fit human subjective perception mechanisms, optimizing the generation results of future AI models.

Overview: RAGoon is a high-level library designed for batched embeddings generation, blazingly-fast web-based RAG, and quantized indexes processing, aiming to enhance language model performance and improve contextually relevant information retrieval. Key features: - Batched embeddings generation - Web-based RAG (Relevance-Augmented Generation) - Quantized indexes processing - High-dimensional vector visualization Tech stack: Jupyter Notebook Why it matters: RAGoon simplifies the process of embedding large datasets, enabling efficient and contextually relevant information retrieval through a web-based interface, making it an essential tool for developers working on NLP and language model projects. Repository: louisbrulenaudet/ragoon GitHub stars: 70 GitHub description: High level library for batched embeddings generation, blazingly-fast web-based RAG and quantized indexes processing ⚡

### Overview This repository contains the implementation of the **ICLR 2025 paper** titled "Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion" by Marco Mistretta, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Andrew David Bagdanov. The paper explores the impact of intra-modal misalignment in contrastively trained VLMs and introduces two optimization-based modality inversion techniques to improve performance in intra-modal tasks. ### Key Features - **Optimization-based Textual Inversion** (OTI): Maps visual features to textual features. - **Optimization-based Visual Inversion** (OVI): Maps textual features to visual features. These techniques are designed to empirically demonstrate the impact of intra-modal misalignment in contrastively trained VLMs. ### Tech Stack The repository is written in Python. ### Why It Matters The paper highlights the common practice of exploiting only one encoder in contrastively trained VLMs, which leads to suboptimal performance in intra-modal tasks. By leveraging modality inversion, the paper shows that approaching intra-modal tasks inter-modally significantly improves performance compared to intra-modal baselines on more than fifteen datasets. Additionally, it demonstrates that approaching a native inter-modal task (e.g., zero-shot image classification) in an intra-modal manner decreases performance, further validating the findings. ### Repository The primary language of the repository is Python, and it has been starred by 70 GitHub users, indicating its popularity and usefulness in the community. ### GitHub

Overview: Gemma4.java is a fast, zero-dependency inference engine for Gemma 4 models written entirely in Java, optimized for performance and ease of use. Key features: - Single-file, no dependencies, based on llama3.java - Supports all Gemma 4 model families: E2B, E4B, 31B, and 26B-A4B (MoE) - Fast GGUF format parser - Supported dtypes/quantizations: F16, BF16, F32, Q4_0, Q4_1, Q4_K, Q5_K, Q6_K, Q8_0 - Matrix-vector kernels using Java's Vector API - CLI with --chat and --prompt modes - Thinking mode control with --think off|on|inline - GraalVM Native Image support - AOT model preloading for instant time-to-first-token Tech stack: Java 21+ Why it matters: Gemma4.java provides a lightweight, high-performance solution for inference tasks, making it ideal for applications requiring fast and efficient model processing without the overhead of external dependencies.

### Overview: Misata is a powerful synthetic data generator for Python that ensures your data meets specific criteria. It allows you to declare the facts your dataset must satisfy, and it builds multi-table data that is verified and deterministic. Misata is particularly useful for generating demo data, seed data, or sales figures. ### Key features: - **Declarative Data Definition**: Write down the facts your dataset must satisfy. - **Multi-Table Support**: Generates data across multiple tables with verified foreign-key integrity. - **Deterministic Output**: Ensures that the generated data is consistent and predictable. - **No Model in the Data Path**: Data is generated directly from your input specifications, without any underlying model. ### Tech stack: Python + MCP server ### Why it matters: Misata is a valuable tool for anyone looking to quickly generate synthetic data that meets specific criteria. Its declarative approach ensures that the data generated is both accurate and consistent, making it perfect for use cases such as demo data, seed data, or sales figures. The Python + MCP server combination makes it easy to integrate and use Misata in various projects.

### Overview: BeatLearning is an open-source project aimed at democratizing the process of automatic beatmap generation for rhythm games. The goal is to create accessible tools and foundation models for developers, players, and enthusiasts, paving the way for a new era of creativity and innovation in rhythm gaming. ### Key Features: - **Automatic Beatmap Generation**: Generates beatmaps for various rhythm games, including OSU, Taiko, and Mania. - **Open Source**: Available on GitHub, with Python as the primary language. - **Community Contributions**: Encourages contributions for improved usability, model integration, and more. - **Model Availability**: Models are available on HuggingFace for further exploration and integration. ### Tech Stack: - Python 3.12 ### Why It Matters: This project aims to make rhythm game beatmap creation more accessible, allowing anyone to create and share their own beatmaps without the need for specialized software or high-level programming skills. By democratizing the process, it opens up new possibilities for creativity and innovation in rhythm gaming. ### Repository: - **Repository**: sedthh/BeatLearning - **GitHub Stars**: 70 - **GitHub Description**: Open Source Generative AI Models for Automatic Rhythm Game Beatmap Generation (for acoustic people) ### How to Use: - **Install Python 3.12**: Ensure you have Python 3.12 installed. - **Create Virtual Environment**: Use `venv` to create a virtual environment. - **Install Required Libraries**: Run `pip3 install -r requirements.txt`

Overview: This project is a deep research server using Google Gemini 2.5 Flash, designed for seamless integration with the Model Context Protocol (MCP) and Google Search Grounding. It aims to provide a simple yet effective implementation of a research AI agent. Key features: - Runs as a Model Context Protocol (MCP) server/tool for seamless agent integration. - Long-context reasoning, structured JSON outputs, and tool use via env flags. - Iterative deep dive with query refinement and result analysis. - Depth & Breadth Control: Tune exploration scope precisely. - Semantic/Recursive Splitting: Token-aware chunking for robust summarization and analysis. - Batching + Caching: Concurrency-limited batched model calls with LRU caches across prompts/results. - Professional Reports: Generates structured Markdown reports. Tech stack: TypeScript Why it matters: This project simplifies the implementation of a deep research agent, making it easy to understand, modify, and extend. It provides a minimal, understandable core with deterministic outputs and clean MCP server entry, ensuring consistent report scaffolding and agent readiness. GitHub stars: 69 Repository: ssdeanx/deep-research-mcp-server