Projects

### Overview: Granite Switch is a platform that allows developers to build AI models by assembling pre-trained adapter functions, making it easy to combine independently developed components into a single model. This approach leverages Activated LoRA technology for efficient deployment, enabling models to be swapped, upgraded, or customized without retraining. ### Key Features: - **Composable**: Combine independently developed adapter functions into one checkpoint, whether IBM's or yours. Swap, upgrade, or customize without retraining. - **Fast**: Built on IBM's Activated LoRA technology for efficient KV cache reuse, low latency, and high inference throughput. - **Accurate**: Task-specific adapter functions can match and even surpass the accuracy of significantly larger generalist models, while requiring only a fraction of the serving cost. - **Inference-ready**: Deploy with vLLM for production or HuggingFace for prototyping. Same checkpoint, no conversion step. ### Why It Matters: Granite Switch simplifies the process of building AI models by providing a library of pre-trained adapter functions that can be combined independently. This approach enables developers to leverage the strengths of existing models while still achieving high performance and efficiency.

Overview: model-compose is a declarative YAML file that allows developers to build and deploy AI systems, models, agents, speech, vision, and live broadcast without writing orchestration code. It simplifies the process by using interchangeable building blocks and allows running the same definition anywhere, including local models and cloud APIs. Key features: - Composable components - Portable definitions - Model-agnostic approach - Stream-native data handling Tech stack: model-compose is built on Python and uses YAML to define the AI system. Why it matters: model-compose simplifies the deployment and orchestration of AI systems, making them more accessible and easier to manage, especially for developers who might not be familiar with orchestration tools or frameworks.

### Overview: This repository collects papers on Large Language Model Enhanced Recommender Systems (LLMERS), focusing on the advancements and applications of LLMERS in the field of recommendation systems. ### Key features: - **Knowledge Enhancement**: Enhances the conventional recommender systems by integrating large language models. - **Interaction Enhancement**: Improves the interaction between users and the system. - **Model Enhancement**: Enhances the recommendation model itself. ### Tech stack: Unknown ### Why it matters: LLMERS, or Large Language Model Enhanced Recommender Systems, has gained significant attention due to its ability to augment conventional recommender systems with large language models, enhancing their performance and relevance in modern recommendation systems.

### Overview: The `manim-generator` project automates the creation of videos using an AI-driven workflow, leveraging the Manim library and a feedback loop involving a Code Writer and Code Reviewer. ### Key features: - **Automatic Video Generation**: Uses an AI-driven workflow to generate Manim videos. - **Code Review and Validation**: Feedback loop with a Code Writer and Code Reviewer. - **Model Comparison**: Uses LiteLLM for comparing different models on the same task. - **Leaderboard**: Manim Bench leaderboard showcasing model performance on full Manim video generation. ### Tech stack: Python ### Why it matters: The project aims to improve video consistency and reduce render failures by leveraging AI-driven feedback and model comparison, making it a valuable tool for anyone looking to automate video creation.

### Overview: A Discord bot designed for chatting with Large Language Models (LLMs) such as LLaMA, Vicuna, Alpaca, or any other LLaMA-based model. It offers a conversational experience and allows users to interact with the bot through `/reply`. ### Key features: - Engaging conversations with LLaMA and its derivatives. - Clear chat history management with `/reset`. - Support for various LLaMA models. - Installation instructions provided for Windows, Linux, and macOS. ### Tech stack: Python ### Why it matters: This bot provides a fun and interactive way to chat with LLaMA models, making the experience more engaging and enjoyable compared to other models.

Overview: AgentGenesis is an open-source web app that provides customizable Gen AI code snippets for seamless integration into applications. Key features: - Copy and paste code snippets - Easy integration into projects - Open-source and no dependencies required Tech stack: TypeScript Why it matters: AgentGenesis simplifies the process of integrating Gen AI code into applications, making it accessible and easy to use for developers.

Overview: 这个项目是一个开放的 CS/AI 知识库，包含基础知识、数据结构与算法、软件开发以及大模型等体系化的笔记。 Key features: - 基础知识 - 数据结构与算法 - 软件开发 - 大模型 Tech stack: Python 3.13, Zensical 0.0.74 Why it matters: 本项目旨在提供计算机与人工智能领域的系统化知识，帮助开发者和学习者提升技能。

### Overview: This repository contains a comprehensive collection of specialized workflow packs for ComfyUI, covering a wide range of AI image and video generation capabilities, including advanced text-to-video workflows, high-quality image generation, 3D generation, video processing, audio integration, and utility tools. ### Key Features: - **CODA-Cosmos-Pack**: Advanced text-to-video generation workflows. - **CogVideo**: Suite of CogVideo implementation workflows. - **cosXL Pack**: SDXL-focused workflows for high-quality image generation. - **DJZ-3D**: 3D generation workflows (SV3Du, TripoSR, Zero123). - **Foda_Flux**: Comprehensive collection including ControlNet implementations, Inpainting workflows, IP Adapter integrations, Upscaling and enhancement tools, Scene creation and editing, and Training tools. - **Donut-Mochi-Video**: Video generation and processing. - **Hoda-Hunyuan-Video**: Text-to-video and video-to-video workflows. - **LoopedMotion2**: Animation loop creation tools. - **Loki-FaceSwap**: Face swapping and manipulation workflows. - **Lyra AudioTools**: Audio processing and integration. - **MODA-MMaudio-Pack**: Multimodal audio-visual workflows. - **Roda-AudioReactive-Pack**: Audio-reactive visual generation. - **DJZ-Nodes-Examples**: Extensive collection of utility workflows including Video effects, Batch processing tools, Prompt manipulation, Image size adjustment, String manipulation tools, and Model merging and comparison workflows. - **

### Overview: This project, `actionow-ai/c2c_img`, allows Claude Code to generate and edit images using Codex's built-in `image_gen` tool, eliminating the need for an `OPENAI_API_KEY`. The plugin seamlessly integrates with your local Codex, making it feel native and intuitive. ### Key Features: - **Image Generation**: Claude Code can generate and edit images directly within the chat interface. - **Native Experience**: You can simply say "draw me a…" and the image appears instantly. - **Customization**: Edit images by specifying options like `-t` for a transparent background or `-s 1200x628` for a specific size. - **Output Formats**: Supports PNG, JPG, and WebP formats, with the output format inferred from the extension. ### Tech Stack: - Python 3.9 or later - The Codex CLI, installed via `npm i -g @openai/codex` with a ChatGPT account ### Why It Matters: This plugin makes Claude Code feel more native by integrating seamlessly with Codex, eliminating the need for an external API key. It's perfect for those who want to leverage the power of Codex without the hassle of managing API keys. ### Repository: - **GitHub Repository**: actionow-ai/c2c_img - **GitHub Stars**: 113 - **Description**: Native-feeling image generation for Claude Code, powered by your local Codex (ChatGPT) image_gen. No API key needed. ### Install: -

### Overview: This repository includes a comprehensive workshop on AI agents, focusing on the development of AI agents using CorePython, CrewAI, and SmolAgent frameworks. The workshop covers building blocks of AI, workflow patterns, and agentic AI in SQL. ### Key Features: - **Part 1: Building Blocks – The Augmented LLM** - Basic LLM Calls - Structured Output - Tool Use - Retrieval - **Part 2: Workflow Patterns for AI Systems** - Prompt Chaining - Routing - Parallelization - **Part 3: Introduction to Agentic Framework (CrewAI)** - CrewAI Framework Step by Step - Agent: Finance Agent App - **Part 4: Introduction to Agentic Framework(SmolAgent)** - SmolAgent Framework Step by Step - Agent: AI News Agent App - **Part 5: Agentic AI in SQL(SmolAgent)** - SmolAgent Framework Step by Step for SQL - Agent: SQL AI Agent Dashboard ### Tech Stack: Jupyter Notebook ### Why it Matters: This workshop provides a practical guide for developers to create AI agents using various frameworks, enhancing their skills in AI development and application.

Overview: Bedrock Deep Research is an application using Amazon Bedrock and LangChain AWS libraries to automate article generation through AI-powered research, content writing, and image generation. Key features: - Automated Research - Structured Content Generation - Interactive Feedback Loop - AI-Generated Imagery Tech stack: Python Why it matters: This project automates the process of generating articles and reports, making it easier and more efficient for researchers and writers. Repository: aws-samples/sample-bedrock-deep-researcher GitHub stars: 118 GitHub description: Researcher Agent to write blog posts/ articles using Amazon Bedrock and websearch.

Overview: Better Agent is a TypeScript framework designed to build typed, durable agent apps using AG-UI event streaming. It provides a composable runtime for various tools, memory, storage, authentication context, interrupts, resumable runs, plugins, and a typed client across different frameworks. Key features: - Typed end-to-end agent apps - Fits into existing stacks - Easy to package and share reusable behavior as plugins - Easy to build durable conversations and resumable runs with interrupted work Tech stack: TypeScript Why it matters: Better Agent offers a typed, durable approach to building agent apps, making it easier to manage and share code across different frameworks.

### Overview: BlaGPT is an experimental playground designed for benchmarking language model architectures, layers, and techniques on smaller datasets. It aims to provide a flexible environment for experimentation and exploration. ### Key Features: - **Technique Documentation**: Contains detailed explanations and code snippets for various techniques implemented in the project. - **Techniques Implemented**: - Multi-token prediction - Weight tying - Grouped query attention - Capping logits - QKV bias - Zero-init projection layer - Post and pre-RMSNorm - Setting base theta to 1_000_000 - Z-loss regularization - KV-Shifting attention - Dilated Attention (LongNet) - Multi-Head Latent Attention - Per token output bias - DyT Norm - Forgetting Transformer (Vanilla and Pro vers) - Multi-Token Attention - Differential Attention - Softpick - Canon Layer - Parallel Transformer Block - Per Layer Token Embedding - PolyNorm - PolyReLU - TOP loss - Simplified RoPe - Gated Attention - ResFormer (Plus) ### Tech Stack: Python ### Why It Matters: BlaGPT serves as a flexible platform for experimenting with various language model architectures and techniques, providing insights into what works and what doesn't. It's designed to be a fun and educational tool for developers interested in improving language model performance.

Overview: Jacobi Forcing is a novel training technique that converts large language models (LLMs) into causal parallel decoders, significantly improving both speed and accuracy in decoding tasks. Key features: - Fast and accurate causal parallel decoding - High tokens-per-forward performance - Significant wall-clock speedup on coding and math tasks - Retains near-AR generation quality Tech stack: Python Why it matters: Jacobi Forcing addresses the limitations of traditional AR decoding by training models to generate high-quality drafts in parallel, leading to substantial improvements in both speed and efficiency while maintaining high-quality output. Repository: hao-ai-lab/JacobiForcing Primary language: Python GitHub stars: 118

### Overview: FabricDiffusion: High-Fidelity Texture Transfer for 3D Garments Generation from In-The-Wild Images This project introduces **FabricDiffusion**, a method for generating high-fidelity 3D garments from in-the-wild images. The key features include: - **Texture Transfer**: Only need datasets that capture the marginal distribution of individual attributes, bypassing the need for datasets that represent joint distributions of multiple attributes. - **Generalizable Token Embeddings**: The learned token embeddings are generalizable across different generative models. ### Key Features: 1. **Texture Transfer**: - **Model**: Uses pre-trained texture models to transfer textures from in-the-wild images. - **Output**: Generates 3D garments with highly realistic textures. 2. **Print Transfer**: - **Model**: Uses pre-trained print models to transfer prints from in-the-wild images. - **Output**: Generates 3D garments with realistic prints. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. - **PyTorch**: Used for training and inference. - **Diffusers**: A library for generating images using diffusion models. - **Transformers**: Used for token embeddings and other transformer-based models. ### Why It Matters: - **Efficiency**: The method reduces the need for large datasets that represent joint distributions of multiple attributes, making it more efficient. - **Generalizability**: The learned token embeddings are generalizable across different generative models, ensuring consistent results across various applications. ### Updates: - **Jan 2

### Overview: SuperKnowa is a powerful framework designed to accelerate the development of Enterprise RAG (Retriever Augmented Generation) pipelines. It leverages Large Language Models (LLMs) to tackle a wide range of Generative AI use cases, making it a versatile tool for organizations looking to integrate AI-driven text generation into their applications. ### Key Features: - **Enterprise-Scale Pipelines:** SuperKnowa is optimized for handling large volumes of data and private knowledge bases, ensuring scalability and efficiency. - **Pluggable Components:** The framework consists of modular components that can be easily assembled, providing flexibility and customization options. - **Real-Time Queries:** The system supports real-time queries, allowing for quick responses to various AI-driven text generation challenges. - **Private Knowledge Base:** SuperKnowa is designed to work with private knowledge bases, making it ideal for organizations with sensitive data. ### Tech Stack: - **Jupyter Notebook:** The primary language used for development and experimentation with the framework. ### Why It Matters: SuperKnowa simplifies the process of building Enterprise RAG pipelines, making it accessible and efficient for developers and businesses looking to integrate AI into their applications. The framework's modular design and real-time capabilities make it a valuable tool for organizations aiming to leverage Generative AI in a scalable and effective manner. ### Repository Contents: - **Indexing Documents:** - **Elastic Search:** A popular open-source search engine. - **Solr:** A distributed search and analytics engine. - **Watson Discovery:** A cloud-based search and discovery service. -

### Overview: MinkLoc++ is a discriminative multimodal descriptor designed for place recognition, re-localization, and loop closure in robotics and autonomous vehicle applications. It leverages a pair of inputs: a 3D point cloud from a LiDAR and an image from a stereo-center camera. The method achieves state-of-the-art performance on standard place recognition benchmarks and identifies a dominating modality problem during training. ### Key Features: - **MinkLoc++**: A multimodal descriptor for place recognition. - **Late Fusion Approach**: Each modality processed separately, then fused in the final part of the pipeline. - **State-of-the-Art Performance**: Achieves the best results on standard place recognition benchmarks. - **Dominating Modality Problem**: Identified when training a multimodal descriptor, where the network focuses on a modality with a larger overfit to the training data, leading to suboptimal performance on the evaluation set. ### Tech Stack: - **Python**: The primary programming language used. - **PyTorch**: Used for deep learning models. - **MinkowskiEngine**: For efficient computation on point clouds. - **Tensorboard**: For visualization and debugging. - **Color Demosaicing**: For improving image quality. ### Why It Matters: - **State-of-the-Art Performance**: MinkLoc++ outperforms existing methods on standard place recognition benchmarks. - **Dominating Modality Problem**: Identifying and mitigating this issue is crucial for improving the robustness of multimodal descriptors. - **Efficient Computation**: Utilizing Mink

### Overview: This repository contains a comprehensive list of papers on data contamination, specifically focusing on large language models. It serves as a valuable resource for researchers and developers interested in evaluating and improving the robustness of their models against potential contamination issues. ### Key features: - **Reactive Approaches**: Papers that propose methods to detect data contamination after it has occurred. - **Preventative Approaches**: Papers that discuss strategies to prevent data contamination before it happens. - **Analysis Papers**: Papers that formally discuss the data contamination problem and provide experimental findings. - **Tool Papers**: Papers that describe systems or software designed to handle data contamination challenges. - **Dataset Papers**: Papers that release datasets specifically targeted at data contamination. - **Survey Papers**: Papers that provide a comprehensive overview of the field. ### Tech stack: The repository uses Markdown for documentation and GitHub Pages for hosting. ### Why it matters: Data contamination is a critical issue in the era of large language models, where models are trained on vast amounts of data that may inadvertently leak test information during training. This list helps researchers and developers identify and mitigate this risk, ensuring more accurate and reliable model evaluations.

Overview: Rise of Machine is a curated directory of AI tools for makers and small businesses, providing a comprehensive catalog of the latest and most relevant products. Key features: - AI tools sorted by category and newest first. - MCP servers from the Model Context Protocol registry. - Skills sourced from their published sources. - Browse any category, open a card for the full details, search across the catalog, and save tools for future reference. Tech stack: Astro. Why it matters: Rise of Machine offers a reliable and up-to-date resource for AI tools, making it essential for founders and SMBs looking to enhance their projects with cutting-edge technology.

Overview: This repository is a historical experiment that replicates Andy Ayrey's "Backrooms" with configurable conversation templates and example transcripts. It supports multiple models including Opus 3, Sonnet 3.5, GPT 4o, o1-preview, and o1-mini, allowing for varied model interactions and setup instructions reflecting the 2024–2025 provider landscape. Key features: - Configurable conversation templates - Example transcripts for reference - Support for multiple models (Opus 3, Sonnet 3.5, GPT 4o, o1-preview, o1-mini) Tech stack: Python Why it matters: This repository preserves the experiment, configuration, and example transcripts rather than serving as a maintained multi-provider library. It provides insights into the models' interactions and setup instructions for adapting to current endpoints. Repository: scottviteri/UniversalBackrooms GitHub stars: 118