Projects

### Overview: A universal, portable installer for Stable Diffusion WebUI Forge, offering two versions: Classic and Neo. This script simplifies the installation process by using UV for package management, making it ultra-fast and fully portable. ### Key features: - **Universal Installer**: Choose between Forge Classic or Forge Neo. - **Ultra-Fast Installation**: Uses UV for package management, reducing install times. - **Fully Portable**: Includes Embedded Python 3.13 and MinGit, no global installation needed. - **Auto-Fixes**: Automatically installs pre-compiled `insightface` wheels for Classic. - **Smart Cleanup**: Automatically removes temporary setup files after installation. ### Tech stack: Batchfile ### Why it matters: This script simplifies the installation and updates process for Stable Diffusion WebUI Forge, making it accessible and easy to use for both beginners and advanced users.

Overview: This repository is a comprehensive collection of resources, tools, papers, and projects focused on ensuring data quality across various data types, including traditional structured and unstructured data, large language model data, multimodal data, and tabular data. It serves as an essential reference for researchers and practitioners in the field of data-centric AI. Key features: - Traditional Data: Papers, Tools & Projects, Data Readiness Assessment - Large Language Model Data: Pretraining Data, Fine-tuning Data, LLM Data Management, Cognition Engineering & Test-Time Scaling - Multimodal Data: Papers, Tools & Projects - Tabular Data: Papers, Tools & Projects - Time Series Data: Papers, Tools & Projects - Graph Data: Papers, Tools & Projects - Data-Centric AI: Surveys, Data Valuation, Data Selection, Benchmarks Tech stack: Python, JavaScript, HTML, Markdown, and various open-source tools and libraries. Why it matters: This repository provides a wealth of resources and insights that are crucial for ensuring the reliability and effectiveness of data in various applications and research projects, making it an invaluable resource for professionals in the field of data-centric AI.

Overview: This repository contains the code for the AI Builders Bootcamp #3, focusing on building AI-powered customer segments using Python, OpenAI's API, and sklearn. Key features: - Lightning Lesson: Build AI-Powered Customer Segments (That Scale) - Session 1: Introduction, Software 1.0 - Session 2: Software 2.0, Data Engineering, Machine Learning - Session 3: Software 3.0, Foundation Models, Prompt Engineering - Session 4: Software 3.0, RAG, Text Embeddings - Session 5: Software 3.0, Fine-tuning Tech stack: Python, OpenAI API, sklearn, Jupyter Notebook Why it matters: This repository provides a comprehensive guide for developers interested in building AI-powered applications, focusing on practical examples and hands-on projects.

Overview: A2A-NT simulates consumer-market negotiations between buyers and sellers using LLM agents, providing full dialogue traces, seller offers, final deal outcomes, budget constraints, and anomaly labels. Key features: - Full dialogue traces - Seller offers extracted - Final deal outcomes recorded - Budget and wholesale constraints applied - Post-run anomaly labels Tech stack: Python Why it matters: This project benchmarks and models agent-to-agent negotiations and transactions in consumer markets, offering valuable insights into consumer behavior and market dynamics through simulated interactions. Repository: ShenzheZhu/A2A-NT Primary language: Python GitHub stars: 32 GitHub description: Official code of "The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets"

Overview: This project presents Audio Lottery, a lightweight speech recognition model that is noise-robust, transferable, and compatible with structured sparsity. The model achieves ultra-lightweight performance while maintaining high accuracy, with the smallest winning ticket containing only 4.4% of the full model's weights. Key features: - Noise-robust speech recognition - Transferable across different datasets - Structured sparsity compatibility Tech stack: Python Why it matters: This project demonstrates the effectiveness of using lottery ticket hypothesis to discover lightweight speech recognition models, showcasing the potential of achieving ultra-lightweight, noise-robust, and transferable models without compromising performance.

### Overview: WhereIsAI is a comprehensive repository that gathers AI companies, products, and tools across various industries. Our goal is to help users find the right AI solution to enhance their work efficiency. ### Key Features: - **Copilot**: A Code Assistant using OpenAI Codex to suggest code in real-time. - **Tabnine**: A Code Assistant that suggests code completions aligned with coding patterns. - **Codewhisperer**: A Code Assistant powered by Amazon CodeWhisperer, helping developers improve productivity by generating code. ### Tech Stack: Multilingual ### Why It Matters: This repository aims to provide a diverse collection of AI solutions, making it easier for developers to find the right tool or product that fits their needs.

Overview: This project is a Gemini Vision & Image Generation MCP server designed specifically for use with Claude Desktop, Claude Code, and Cursor. It supports Nano Banana 2 and Nano Banana Pro models, offering image generation, editing, session consistency, and runtime model switching. Key features: - Image Generation - Image Editing - Session Consistency - Runtime Model Switching Tech stack: JavaScript Why it matters: This MCP server streamlines Gemini's image capabilities, making it easier to integrate and utilize in Claude's desktop and code environments, enhancing both image creation and editing functionalities. Repository: YCSE/nanobanana-mcp Primary language: JavaScript GitHub stars: 32 GitHub description: Gemini Vision & Image Generation MCP for Claude Desktop and Claude Code README excerpt: This README provides an overview of the NanoBanana MCP project, including its features, quick start instructions, and tools for integration with Claude Desktop, Claude Code, and Cursor.

Overview: afdEP-comm is an expert-parallel communication system for Attention–FFN Disaggregation (AFD) built on DeepEP V2, designed to optimize performance and reduce latency. Key features: - Expert-parallel communication between Attention GPUs and FFN GPUs. - Vendor of DeepEP V2's kernels and specialized them for the two roles. - Routing, quantization, expert compute, and serving scheduling remain with the caller. - Benchmarks run against FastAFD's DeepEP M2N, showcasing improvements in performance and efficiency. Tech stack: Python Why it matters: By optimizing the communication between different components of the AFD system, afdEP-comm significantly reduces latency and improves overall performance, making it a valuable tool for developers working on high-performance computing and machine learning applications.

### Overview ReAG is a Reasoning-Augmented Multimodal Retrieval-Augmented approach designed for Knowledge-based Visual Question Answering (KB-VQA). It addresses the issue of noisy or irrelevant passages often retrieved by standard retrieval-augmented methods, by combining coarse- and fine-grained retrieval with a critic model that filters out low-quality passages before generating the answer. The model is trained using a multi-stage strategy: a supervised fine-tuning as a cold start, followed by reinforcement learning to promote explicit reasoning grounded in retrieved evidence. ReAG significantly outperforms prior methods on Encyclopedic-VQA and InfoSeek. ### Key Features - **ReAG Generator 3B**: A model checkpoint for the ReAG generator with 3 billion parameters. - **ReAG Generator 7B**: A model checkpoint for the ReAG generator with 7 billion parameters. - **ReAG Critic**: A model checkpoint for the ReAG critic. ### Tech Stack The repository uses Python for the inference pipeline and hosts the model checkpoints on Hugging Face. ### Why It Matters ReAG significantly improves the quality of answers in Knowledge-based Visual Question Answering tasks, particularly on Encyclopedic-VQA and InfoSeek benchmarks, outperforming previous methods. ### Repository The repository is hosted on GitHub at [aimagelab/ReAG](https://github.com/aimagelab/ReAG). It includes the inference pipeline for ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering. The checkpoints are available on Hugging Face

Overview: Raven is a Python-based tool that generates unique images by combining images found on Google based on the input text. Key features: - Generates unique images - Uses Google to find images - Needs to download images Tech stack: Python Why it matters: Raven simplifies the process of generating unique images with text, making it a useful tool for various creative projects or educational purposes. Repository: billythegoat356/Raven GitHub stars: 32 GitHub description: Raven is a tool written in Python3 allowing you to generate an unique image with some text.

Overview: This project is a TensorFlow Lite for Microcontrollers demo specifically designed for the M5Stack, showcasing facial expression changes based on speech input. Key features: - Uses TensorFlow Lite for Microcontrollers - Demonstrates speech recognition and facial expression changes - Compatible with M5Stack FIRE and M5Stack M5GO Tech stack: Arduino-based M5Stack library on PlatformIO Why it matters: This project provides a practical demonstration of speech recognition capabilities on microcontrollers, showcasing the integration of TensorFlow Lite with microcontroller platforms like M5Stack, which can be useful for various applications such as smart home devices or voice-controlled interfaces.

Overview: The RAG Coding Assistant is an AI tool designed to generate working code in LangChain Expression Language (LCEL) for developing new generative AI prototypes. Key features: - Outputs working code in LCEL - Uses a custom retrieval-augmented generation technique - Retrieves relevant examples from a Weaviate vector database - Provides context for user requests Tech stack: Jupyter Notebook Why it matters: This tool simplifies the process of creating new AI prototypes by providing accurate and relevant examples directly in LCEL, making it easier for developers to start building their projects quickly.

Overview: This comprehensive guide teaches you how to build production-grade distributed applications using Ray, a unified framework for scaling AI and Python applications. Whether you're training large language models, processing terabytes of data, or deploying ML models at scale, Ray simplifies distributed computing by abstracting away the complexity of cluster management and parallelization. Key features: - Distributed training - Multimodal data processing - Inference at scale - Reinforcement learning Tech stack: Python, PyTorch, Ray Why it matters: This guide is essential for anyone looking to leverage Ray for production workloads at scale, making it easier to build scalable distributed applications with Python and AI. Repository: debnsuma/ray-for-developers GitHub stars: 32 GitHub description: A comprehensive hands-on guide to building production-grade distributed applications with Ray - from distributed training and multimodal data processing to inference and reinforcement learning.

Overview: Diffuse is a decentralized peer-to-peer system for large language model inference, where your prompt never leaves your device in clear text. It is designed for privacy and security, aiming to replace cloud-based AI with a network of ordinary machines. Key features: - Decentralized peer-to-peer architecture - Your prompt never leaves your device in clear text - No servers, no surveillance, no prompt logging - Built with Rust and PyTorch Tech stack: Rust and PyTorch Why it matters: Diffuse challenges the traditional cloud-based AI model by providing a decentralized and secure alternative, ensuring privacy and control over AI interactions.

Overview: Graphlit simplifies and accelerates AI and data development by providing a managed platform for unstructured data and LLM integration. Key features: - API-first approach - Graph-based knowledge graph - Multi-modal RAG (Retrieval-Augmented Generation) - Model-agnostic integration with OpenAI, Anthropic, Meta, and others - Managed cloud-native platform - Built-in multi-tenancy, semantic search, storage, and workflow automation Tech stack: Graphlit uses LangChain, Pinecone, S3, and other tools for unstructured data ETL pipelines. Why it matters: Graphlit simplifies the development process by handling data and infrastructure, allowing developers to focus on their core applications without the complexity of integrating multiple components. Repository: graphlit/graphlit GitHub stars: 32 GitHub description: Graphlit Platform README excerpt:

### Overview: SonicBoom is a compiler-first PyTorch alternative designed for performance-critical AI workloads, built entirely in C++ and CUDA. It leverages proven kernels to optimize performance, making it ideal for local LLM inference on consumer GPUs. ### Key features: - **Compiler Optimization:** SonicBoom utilizes a stable C ABI to expose everything through a stable C ABI, ensuring performance parity with other models. - **Performance:** On a desktop RTX 3050, SonicBoom decodes 14-19% faster than Llama.cpp, using the same quantized kernels. It is 1.19× faster on Gemma 4 E4B-it Q3_K_M and 1.14× faster on Qwen2.5-0.5B-Instruct Q4_K_M, with a 13% prefill advantage on Gemma 4 and 28% on Qwen2.5. - **Language Models:** SonicBoom targets language, vision, and training models, with benchmarks against Llama.cpp on four families: Gemma 4 E4B-it (Q3_K_M), Qwen2.5-0.5B-Instruct (Q4_K_M), DeepSeek-R1-Distill-Qwen-1.5B (Q4_K_M), and Qwen2.5-7B-Instruct-1M (Q4_K_M). - **Sampling and Constrained Decoding:** Seeded temperature, top-k, top-p, and min-p penalties on the C

### Overview: Envision is a lightweight inference engine for FLUX2 and ZImage models written in pure Pascal. It runs natively in Delphi, FPC, or Lazarus IDEs, eliminating the need for external dependencies. Envision parses .Safetensors and .JSON files, runs out-of-the-box, and utilizes lazy loading and disk memory mapping for efficient memory usage. It supports cross-platform compatibility and supports Windows, Linux, and macOS with automatic binding to OpenBLAS or the Accelerate Framework for performance improvements. ### Key Features: - **Zero Runtime**: The binary can be shipped to any machine without requiring a Pascal compiler or DLLs. - **Safety & Readability**: Pascal's strong typing and modularity reduce bugs in tensor manipulation. - **Cross-Platform**: Works on Windows, Linux, and macOS with minimal changes. - **Concise Executable**: The output executable is self-dependent, less than 4MB (Windows). - **CPU Performance**: Hand-written Intel x86_64 assembly code to squeeze out the best CPU performance possible. - **Performance Improvements**: On macOS, automatically utilizes the Accelerate Framework for ~3x speed-up. - **Image Edit Support**: Supports FLUX2 image edit functionalities, including reference image to image and multi-ref image to image. ### Tech Stack: - Pascal ### Why It Matters: - **Zero Runtime**: The binary can be shipped to any machine without requiring a Pascal compiler or DLLs. - **Safety & Readability**: Pascal's strong typing and modularity reduce bugs in tensor

Overview: TULIP, or Token-length Upgraded CLIP, is a method designed to enhance the caption length of CLIP-like models, enabling them to better understand and generate long captions. This project is based on the accepted paper at ICLR 2025, addressing the challenge of long caption understanding. Key features: - TULIP injects relative positional encodings into contrastive vision-language models to handle long image captions. - It provides a seamless integration with CLIP-like models. - It achieves state-of-the-art results on long-caption understanding tasks. Tech stack: Python Why it matters: TULIP is crucial for improving the performance of models that need to generate or understand long captions, making it particularly valuable in applications requiring detailed and comprehensive descriptions of images or text. Repository: ivonajdenkoska/tulip GitHub stars: 32

Overview: This project is an official implementation of the NeurIPS 2025 paper "Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning". It accelerates inference of reasoning language models by selectively pruning KV cache entries while preserving reasoning quality. Key features: - Efficient Inference for Reasoning LLMs - Training-Free Compression - Semantic-Aware Pruning - Significant Throughput Gains - Memory Usage Reduction Tech stack: Python Why it matters: This method improves throughput and reduces memory usage with minimal accuracy drop, making it suitable for accelerating inference of reasoning language models. Repository: jiwonsong-dev/ReasoningPathCompression GitHub stars: 32 GitHub description: Official implementation of the NeurIPS 2025 paper "Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning".

Overview: Voxtral is a cutting-edge speech understanding model designed for high-quality transcription and audio analysis, optimized for long-form context, multilingual capabilities, and natural language understanding. The demo interface allows users to evaluate Voxtral's performance and explore its applications for transcription and deeper analysis. Key features: - Speech Transcription: Generates high-accuracy transcriptions from audio, supporting up to 32k tokens. - Audio Understanding: Answers questions or generates structured summaries directly from speech. - Multilingual: Supports automatic language detection and state-of-the-art performance in widely used languages. - Built-in Q&A and Summarization: Directly asks questions or generates summaries without separate models. - Function-Calling from Voice: Triggers backend functions based on spoken user intents. - Affordable and Open: Bridges open-source ASR systems and proprietary APIs, delivering production-ready speech understanding at a fraction of the cost. Tech stack: Python Why it matters: This demo interface provides a user-friendly way to evaluate Voxtral's performance and explore its applications, making it accessible for developers and researchers looking to integrate speech understanding into their projects.