Projects

Overview: Wav2Keyword is a keyword spotting (KWS) model based on Wav2Vec 2.0, showcasing state-of-the-art performance in the Google Speech Commands datasets V1 and V2. Key features: - State-of-the-art keyword spotting using Wav2Vec 2.0 backbone - Pretrained Wav2Vec 2.0 base model available in the model directory - Supports both V1 and V2 datasets - Benchmarks can be run to measure accuracy on a given dataset Tech stack: Python Why it matters: This model provides a powerful and efficient solution for keyword spotting in speech data, demonstrating significant improvements over traditional methods in recognizing keywords in speech commands datasets.

Overview: This repository provides a one-command installation script for vLLM on NVIDIA DGX Spark systems with Blackwell GB10 GPUs, focusing on the sm_121 architecture. It includes CUDA 13.0 support with Blackwell-specific optimizations, Triton 3.5.0 from the main branch, and PyTorch 2.9.0 with CUDA 13.0 bindings. Key features: - One-command installation script for vLLM on DGX Spark with Blackwell GB10 GPUs - Handles all complexities of building vLLM on the DGX Spark platform - CUDA 13.0 support with Blackwell-specific optimizations - Critical fixes for SM100/SM120 MOE kernel compilation - Triton 3.5.0 from the main branch (required for sm_121a support) - PyTorch 2.9.0 with CUDA 13.0 bindings - All necessary build fixes and workarounds Tech stack: Shell Why it matters: This script simplifies the installation process for vLLM on DGX Spark systems, ensuring compatibility with the Blackwell architecture and providing a stable and optimized environment for model training and inference. Repository: eelbaz/dgx-spark-vllm-setup Primary language: Shell GitHub stars: 110

### Overview: Claude is integrated with the Claude Agent Skills to facilitate image generation, offering a range of tools from a zero-cost code-based design engine to a real diffusion model on Cloudflare. The AI Storybook pipeline turns plain-English stories into illustrated, narrated HTML books, ensuring Claude handles prompt engineering and orchestration. ### Key Features: - **Code-Based Design Engine**: Renders designs with minimal AI model usage, suitable for posters, quotes, and story/reel covers. - **Three.js 3D Renderer**: Produces 3D product shots and rendered scenes, ideal for wallpapers and thumbnail backgrounds. - **Cloudflare Workers AI (Flux)**: Offers a hosted diffusion model for more realistic and detailed images. ### Tech Stack: HTML ### Why it Matters: Claude seamlessly integrates with the Claude Agent Skills to provide a versatile and cost-effective solution for generating images, from simple designs to complex 3D scenes. The AI Storybook pipeline ensures that every step of the process—from prompt creation to final output—is handled by Claude, making it a powerful tool for anyone looking to create illustrated, narrated HTML books.

Overview: M\* (pronounced "M-star") is a high-performance, universal serving system designed for any-to-any composite multimodal models. It matches or beats state-of-the-art inference systems across various modalities. Key features: - Fast per-component fast paths - Flexible abstraction mirroring the model - One small Python file per model - YAML file for component mapping - Python SDK, OpenAI-compatible API, and native streaming endpoint - Supports text, image, audio, video, and robot actions Tech stack: Python 3.12 Why it matters: M\* is a game-changer for serving composite multimodal models, offering unmatched performance across various modalities and models. Repository: mstar-project/mstar GitHub stars: 107

Overview: This library offers a collection of advanced models for various Natural Language Processing (NLP) tasks, including classification, automatic speech recognition, text summarization, machine translation, and more. It simplifies the training and inference of these models, making them accessible for developers. Key features: - Text-based Graph Convolution Networks (GCN) - Bidirectional Encoder Representations from Transformers (BERT) - XLNet - Graph Attention Network (GAT) - ALBERT - XLMRoBERTa - Graph Isomorphism Network (GIN) Tech stack: Python Why it matters: This toolkit provides a comprehensive set of models for NLP tasks, making it easier for developers to implement advanced NLP solutions without the need for extensive model development and training. Repository: plkmo/NLP_Toolkit GitHub stars: 110 GitHub description: Library of state-of-the-art models (PyTorch) for NLP tasks GitHub URL: https://github.com/plkmo/NLP_Toolkit

Overview: FLEX is a generative model designed for full-body grasping, leveraging pre-trained prior models for full-body pose, right-hand grasping, and pose-ground relation. This implementation can be used to generate full-body avatars grasping objects in various scenes. Key features: - Runs FLEX on arbitrary objects in arbitrary scenes - Supports pre-computed object centering and BPS representation for GRAB dataset objects Tech stack: Python, PyTorch, MANO, SMPLX, chamfer_distance Why it matters: FLEX provides a practical solution for full-body grasping tasks, making it valuable for applications requiring realistic full-body avatar interactions in virtual environments.

### Overview: MMChat is a large-scale dialogue dataset containing image-grounded dialogues in Chinese, designed for the LREC2022 paper "MMChat: Multi-Modal Chat Dataset on Social Media". The dataset features 198.82K images and includes 314.13K utterances across 120.84K sessions, with an average of 2.599 utterances per session and 2.791 images per session. ### Key Features: - **Image-Grounded Dialogues**: Each dialogue is associated with one or more images (maximum 9 images per dialogue). - **Quality Assurance**: Various strategies ensure the quality of the dialogues. - **Data Availability**: The dataset is available through the Hugging Face datasets library, making it easy to access and use. ### Tech Stack: Python ### Why it Matters: The MMChat dataset provides a rich resource for researchers and developers interested in natural language processing, specifically in the areas of dialogue understanding and multimodal interaction. Its structured nature and the inclusion of images make it particularly valuable for training and evaluating models that can handle both text and visual inputs.

Overview: Faster Whisper is a local Speech-to-Text engine for Home Assistant, allowing users to create their own personal voice assistant. Key features: - Custom integration with a local Speech-to-Text engine - Built-in support for Home Assistant - Uses a local, lightweight engine for improved performance - Compatible with various models and configurations Tech stack: Python Why it matters: This project offers a practical solution for users who want to integrate a voice assistant into their Home Assistant setup without relying on the official Whisper App, providing a more efficient and resource-efficient alternative.

### Overview: This repository is an extensive collection of papers and resources focused on spoken language identification (LID). It serves as a valuable resource for researchers and practitioners looking to enhance their understanding and skills in this field. ### Key Features: - **Papers**: Over 100 papers covering various aspects of LID, including conference proceedings, workshops, and challenges. - **Datasets**: Includes datasets such as NIST Language Recognition Evaluation, OpenSLR, Common Voice, OLR dataset, Voxlingual107, and FLEURS. - **Workshops/Challenges**: Features workshops and challenges related to LID, including the OLR Challenge, NIST LRE, and 1st WSTCSMC. - **Toolkits**: Provides various toolkits for LID, such as Kaldi, hyperion, ASV-Subtools, SpeechBrain, lidbox, WeNet, ESPNet, s3prl, and Hugging Face. ### Tech Stack: - **Python**: The primary programming language used for implementing and testing LID models. ### Why It Matters: This repository is crucial for researchers and practitioners looking to stay updated with the latest advancements in spoken language identification. It provides a comprehensive overview of the field, including practical tools and datasets, making it an essential resource for anyone aiming to improve their skills in this area. ### GitHub Stars: - **GitHub Stars**: 109 ### Repository: Repository: Lhx94As/Awesome-Spoken-Language-Identification

Overview: This project is a custom interior design pipeline API that utilizes the Realistic Vision V3.0 inpainting pipeline, segmentation, and MLSD ControlNets to generate realistic interior designs based on text prompts. Key features: - Realistic Vision V3.0 inpainting pipeline - Segmentation and MLSD ControlNets integration - Dockerized API for serving the model Tech stack: Python Why it matters: This project provides a practical tool for interior designers and enthusiasts to create visually appealing and realistic interior designs using text prompts, making it accessible and efficient for design projects.

### Overview: Bedrock Agentic ChatBot is a personal private chatbot powered by Amazon Bedrock LLMs, featuring a data analytics feature that provides isolated serverless compute on for data analysis. The chatbot offers a web-based intuitive user interface accessible from any device, handles multi-modal document upload and chat, and maintains privacy of conversations. ### Key features: - **Agentic Orchestration**: A Strands `Agent` owns every conversation. Tools are opt-in per session from the sidebar; the orchestrator decides when to invoke them and relays sub-agent results. - **Broad Model Catalog, Three Engines**: Supports Bedrock Runtime models via the Converse API and Bedrock Mantle OpenAI-compatible models via the Responses API and Chat Completions API. - **Advanced Data Analytics**: A data-analysis sub-agent writes and runs Python against your attached datasets in a sandbox, iterates on errors, validates its own charts, and returns a written analysis with interactive Plotly charts. Two runtimes, selected by a slider: `python` (serverless, stateful IPython kernel) and `pyspark` (Amazon Athena for Apache Spark). - **Document Generator & Editor**: Sub-agent produces and edits Word/PowerPoint/PDF/Excel files in the Code Interpreter sandbox, verifies them structurally, and delivers them as presigned download links in an artifacts expander. - **Research Workers**: The orchestrator fans self-contained research briefs out to parallel ephemeral worker agents, keeping its own context clean. Progress is shown per worker; citations are aggregated into the Sources

Overview: MCP2Lambda is a tool that allows Large Language Models (LLMs) to run AWS Lambda functions as generative AI tools without requiring any code changes. It bridges the gap between the model and the AWS Lambda environment, enabling access to private resources and external services. Key features: - Enables LLMs to invoke AWS Lambda functions directly. - Provides a bridge between MCP clients and Lambda functions. - Supports both pre-discovery and generic mode strategies for Lambda function discovery and invocation. - Integrates seamlessly with AWS Lambda, VPCs, and other AWS services. - Uses the Model Context Protocol (MCP) for communication. Tech stack: Python Why it matters: This approach enhances the capabilities of LLMs by allowing them to interact with AWS Lambda functions, enabling access to private data, external services, and specialized processing capabilities. It simplifies integration and security, making it easier for models to leverage AWS resources without the need for public network access.

### Overview: WindTunnel is a Python-based market simulation engine designed to evaluate trading strategies and market impact under synthetic yet realistic conditions. It combines macro regime generation, high-frequency order flow simulation, LOB microstructure modeling, and plug-in execution strategies, making it a versatile tool for both controlled experimentation and generative scenario construction. ### Key Features: - **Modular Design**: WindTunnel is built on a modular architecture, allowing for easy customization and extension. - **Macro Regime Generation**: Simulates different market regimes to test strategies under various conditions. - **High-Frequency Order Flow Simulation**: Models the behavior of order flow at high frequencies to enhance the realism of the simulation. - **LOB Microstructure Modeling**: Captures the dynamics of the order book to provide a more accurate representation of market behavior. - **Plug-In Execution Strategies**: Supports execution strategies such as VWAP, TWAP, and Dynamic Stop Loss, enabling a wide range of trading behaviors. - **Controlled Experimentation and Generative Scenarios**: Allows users to experiment with different market conditions and scenarios. ### Tech Stack: Python ### Why it Matters: WindTunnel is essential for developers and traders looking to evaluate and optimize trading strategies under various market conditions. Its ability to simulate realistic market scenarios and execute strategies in a controlled environment makes it a valuable tool for both research and development purposes. ### Repository: jialuechen/windtunnel ### GitHub Stars: 109 ### GitHub Description: Python Market Simulation Engine Built on top of Generative AI

Overview: This repository contains the source code for the NeurIPS 2021 and TPAMI 2024 papers titled "Recognizing Vector Graphics without Rasterization" and "Hierarchically Recognizing Vector Graphics and A New Chart-based Vector Graphics Dataset". The project introduces YOLaT and YOLaT++, two powerful models designed to recognize vector graphics without rasterization. Key features: - YOLaT & YOLaT++ - Floorplans and Diagrams datasets for training and inference - Hierarchical structure for VGs - Position-aware enhancement for differentiating similar primitives Tech stack: Python Why it matters: This project addresses the challenge of recognizing vector graphics without rasterization, which is crucial for applications that require efficient and accurate recognition of vector graphics, such as in document analysis or image processing tasks. Repository: microsoft/YOLaT-VectorGraphicsRecognition GitHub stars: 109

Overview: Lexoid is an open-source universal adapter for large language models designed to turn messy real-world data into clean, agent-ready context. Key features: - `parse()`: High-accuracy document and webpage parsing across static, local LLM, and frontier LLM pipelines. - `browse()`: Goal-driven, evidence-grounded browser automation for live web workflows. Tech stack: Python Why it matters: Lexoid simplifies the process of extracting useful information from documents and websites, making it easier to automate tasks without the need for custom pipelines.

### Overview: ggufy is a lightweight and efficient tool designed for converting tensor formats. It supports both CLI and GUI flavors and is optimized for fast and memory-efficient operations. The tool is capable of converting between various formats, including safetensors and gguf quantizations, and supports sensitivity-aware quantization files for specific architectures. ### Key Features: - **CLI and GUI Flavors**: Available for Linux, Windows, and macOS (arm64 and x86_64). - **Fast and Memory-Efficient**: Optimized for performance and resource usage. - **Safetensors and GGUF Support**: Converts between safetensors and gguf formats. - **Sensitivity-Aware Quantization**: Handles conversion with sensitivity-aware quantization files. - **Image and Video Models**: Supports conversion for image and video diffusion models. - **Limited LLM Support**: Currently targets a limited subset of language models. ### Tech Stack: Zig ggufy is written in Zig, a language known for its efficiency and performance. ### Why It Matters: ggufy simplifies the process of converting tensor formats, making it easier for developers to work with different models and formats. Its fast and memory-efficient nature makes it suitable for real-time applications and resource-constrained environments. ### GitHub Stars: 109 The repository has gained significant traction among the developer community, indicating its usefulness and popularity. ### Download: You can download pre-built executables from the releases page. ### Join the Community: Join our Discord channel to discuss ggufy, AI, and

Overview: This repository contains the PyTorch implementation of LDMSeg, a simple latent diffusion approach for panoptic segmentation and mask inpainting. The provided code includes both training and evaluation. Key features: - Training Auto-Encoder - Training LDM (Latent Diffusion Model) conditioned on RGB images Tech stack: Python Why it matters: LDMSeg simplifies panoptic segmentation by relying on Stable Diffusion, avoiding specialized architectures, complex loss functions, and post-processing methods. It achieves promising results for global mask inpainting on scene-centric datasets. GitHub stars: 109

### Overview: CHATS is a cutting-edge framework that combines human preference alignment with test-time sampling to generate high-quality text-to-image results. It leverages a unified model that models both preferred and dispreferred distributions, ensuring superior alignment, fidelity, and aesthetic consistency in text-to-image generation. ### Key Features: - **Human-Aligned Fine-Tuning with CFG Integration**: CHATS integrates human preference alignment with classifier-free guidance sampling, creating a unified framework. - **Proxy-Prompt Sampling**: CHATS uses proxy-prompt sampling to leverage useful signals from both preferred and dispreferred distributions at test time, leading to superior results. - **Data Efficiency**: CHATS achieves state-of-the-art results across benchmarks with minimal fine-tuning effort on a small, high-quality dataset. - **Plug-and-Play**: CHATS is compatible with any diffusion backbone and existing guidance methods, making it highly versatile. ### Tech Stack: - **Framework**: PyTorch - **Model**: Stable Diffusion XL (SDXL) and Stable Diffusion 1.5 (SD1.5) ### Why It Matters: CHATS is a significant advancement in the field of text-to-image generation, offering unparalleled quality and efficiency. Its ability to align human preferences with the output of AI models makes it particularly valuable for applications requiring high-quality, contextually relevant images. ### Repository: - **Repository**: [ATH-MaaS/CHATS](https://github.com/AIDC-AI/CHATS) - **Stars**: 109 - **Description**: Combining Human-Aligned Optimization and Test-Time

### Overview PasLLM is a high-performance Large Language Model inference engine written in pure Object Pascal. It supports multiple model architectures and advanced 4-bit quantization formats for efficient model deployment. ### Key Features - **Pure Object Pascal**: No Python or external dependencies for inference. - **Cross-Platform**: Compatible with Delphi ≥11.2 and FreePascal ≥3.3.1. - **Multiple Architectures**: Support for Llama, Qwen, Phi, Gemma, Mixtral, and more. - **Advanced Quantization**: Custom Q4*NL formats (Q40NL, Q41NL, Q42NL, Q43NL) with superior tail reconstruction. - **Optimized Performance**: Native Pascal implementation with platform-specific optimizations. - **CLI and GUI**: Both command-line interface and visual applications (FMX, VCL, LCL). ### Quantization Formats PasLLM implements several custom 4-bit and 8-bit quantization formats designed for optimal quality/size tradeoff: - **Q40NL**: 4.5 bits/weight with non-linear decode (often better than Q40). - **Q41NL**: Alternative non-linearity with increased tail emphasis. - **Q42NL**: Enhanced variant with improved reconstruction. - **Q43NL**: Advanced format with multiple optimization methods (gradient, coarse-fine, grid). - **Q40**: 4 bits/weight standard quantization (matches llama.cpp Q4_0 quality). - **Q80

Overview: VectorETL simplifies the process of converting and embedding data for use in vector databases and Generative AI applications. Key features: - Modular architecture supporting multiple data sources, embedding models, and vector databases. - Batch processing for efficient handling of large datasets. - Configurable chunking and overlapping for text data. - Easy integration of new data sources, embedding models, and vector databases. Tech stack: Python Why it matters: VectorETL streamlines the creation and management of vector search systems, making it easier for developers and data scientists to build and scale applications that require semantic search, recommendation systems, or other vector-based operations.