Projects

### Overview: EnterpriseRAG Bench provides a comprehensive dataset of over 500,000 company internal documents and 500 questions, designed to benchmark and fine-tune RAG systems on realistic enterprise data. The dataset simulates a fictional company called "Redwood Inference" and covers various business activities such as planning, engineering, documentation, and internal discussions. It includes questions categorized into 10 different types, ranging from simple to complex, to test the robustness and accuracy of RAG models. ### Key features: - **500,000 Documents and 500 Questions**: Comprehensive dataset for benchmarking RAG systems. - **10 Categories of Questions**: Covering a wide range of business activities and question types. - **Industry-Specific Data**: Simulated from various sources including Slack, Gmail, Linear, Google Drive, Hubspot, Fireflies, GitHub, and Jira. - **Ground Truth Documents**: Provided for most questions to ensure accurate evaluation. ### Tech stack: - **Python**: Used for data preprocessing and model evaluation. - **Hugging Face Transformers**: For fine-tuning and evaluation of RAG models. ### Why it matters: - **Benchmarking RAG Systems**: Helps developers evaluate and improve their RAG models on real-world enterprise data. - **Realistic Testing Environment**: Provides a practical environment to test the performance of RAG systems on a wide range of business scenarios. - **Industry-Specific Data**: Simulates data from different industries and scales, making the dataset more relevant and useful for developers.

### Overview: Hierarchical Attention Networks for Document Classification This project implements Hierarchical Attention Networks (HANs) for document classification using PyTorch. The model learns to identify and classify documents based on their hierarchical structure, combining words into sentences and sentences into documents. This approach leverages the structure of documents to enhance classification performance. ### Key Features: - Hierarchical Attention Networks (HANs) - PyTorch implementation - Joint optimization of word and sentence attention models - Padding with zeros for minibatches - Support for gradient masking in PyTorch (coming soon) ### Tech Stack: PyTorch This project is implemented using PyTorch, a popular deep learning framework. ### Why It Matters: Hierarchical Attention Networks offer a novel way to classify documents by leveraging their hierarchical structure. This method can significantly improve classification accuracy, especially in scenarios where the document structure is informative. The implementation in PyTorch allows for easy integration and experimentation with this approach. ### Repository: EdGENetworks/attention-networks-for-classification ### GitHub Stars: 606 ### GitHub Description: Hierarchical Attention Networks for Document Classification in PyTorch ### Primary Language: Jupyter Notebook The project includes an example of a trained model on the IMDB movie review dataset, demonstrating the model's effectiveness. The preprocessed data is available for download, and the training loss plot shows the model's performance on the dataset.

Overview: Yet Another Language Model (yalm) is an implementation of an LLM inference system in C++/CUDA, designed primarily for educational purposes in performance engineering and LLM inference. The project emphasizes documentation, scientific understanding, and readability. Key features: - yalm is an educational tool for performance engineering and LLM inference. - It uses no external libraries except for I/O. - The codebase is intended for educational and research purposes, not for production use. Tech stack: C++/CUDA Why it matters: yalm serves as a practical demonstration of LLM inference implementation, focusing on performance and optimization techniques. It provides insights into how to efficiently implement large language models using C++ and CUDA. Repository: andrewkchan/yalm GitHub stars: 604 GitHub description: yalm is an LLM inference implementation in C++/CUDA, using no libraries except for I/O. It is intended as an educational exercise in performance engineering and LLM inference implementation. The codebase emphasizes documentation, scientific understanding, and readability. It is not meant for production use. See my blog post [Fast LLM Inference From Scratch](https://andrewkchan.dev/posts/yalm.html) for more details.

### Overview: Chainloop is an open-source infrastructure designed to collect and enforce policies across the software delivery process. It ensures that every AI coding session, pull request, build, SBOM, scan, and test report is signed and timestamped, linking back to the commit and release. This creates a trusted graph of what your agents and pipelines did, allowing you to check against policies as code and create verification loops. ### Key Features: - **Define Guardrails**: Define what good looks like as code. - **Collect Signals**: Gather data from every part of the software delivery process. - **Enforce Continuously**: Policies are enforced continuously, ensuring compliance. - **Signed Records**: Every record is signed and timestamped, providing a chain of custody. - **Built on Sigstore and in-toto**: Uses provenance and security scorecards for enhanced security and compliance. ### Tech Stack: - **Go**: The primary language used for development. - **Apache 2.0**: The open-source license for Chainloop. ### Why It Matters: Chainloop helps teams ensure that their software delivery is aligned with defined policies, providing a trusted and secure way to manage and enforce compliance across the entire development lifecycle. ### Repository: - **Repository**: `chainloop-dev/chainloop` - **Stars**: 606 - **Description**: Agentic SDLC infrastructure. Define guardrails. Collect signals. Enforce continuously. Signed records of what your AI coding agents and CI/CD pipelines did, checked against policies as code, with the loop closed in CI/CD. Built on

### Overview: Asqav SDK: Python + TypeScript Asqav is a Python and TypeScript SDK designed for verifiable evidence of AI agent actions. It provides signed receipts, policy enforcement, and audit trails, ensuring compliance without requiring an account or trust. The SDK works seamlessly with LangChain, CrewAI, and MCP, making it ideal for secure AI development and deployment. ### Key Features: - **Signed Receipts**: Every agent action is accompanied by a signed receipt, ensuring compliance and auditability. - **Policy Enforcement**: Policies can be enforced through the SDK, ensuring that actions adhere to predefined rules. - **Audit Trails**: Comprehensive audit trails are maintained, allowing anyone to verify the actions performed by the AI agent. - **Zero-Native Dependencies**: The SDK is completely self-contained, with cryptography running on the server side, ensuring security and reliability. - **Open Source**: The source code is available for review, and the SDK is licensed under the Elastic License 2.0 and Apache-2.0 for commercial use. ### Tech Stack: - **Python**: The SDK is built using Python, leveraging its robust ecosystem and ease of integration. - **TypeScript**: TypeScript is used for the TypeScript SDK, providing type safety and better code readability. ### Why It Matters: - **Compliance and Trust**: Asqav ensures that AI actions are traceable and verifiable, reducing the risk of fraud and ensuring trust in the AI ecosystem. - **Security**: The SDK's server-side cryptography ensures that sensitive information is protected, even when running on the client side. - **

### Overview: This is an official implementation of CvT: Introducing Convolutions to Vision Transformers. We present a new architecture, named Convolutional vision Transformers (CvT), that improves Vision Transformers (ViT) in performance and efficiency by introducing convolutions into ViT to yield the best of both designs. This approach achieves state-of-the-art performance over other Vision Transformers and ResNets on ImageNet-1k, with fewer parameters and lower FLOPs. Performance gains are maintained when pretrained on larger datasets and fine-tuned to downstream tasks. Pre-trained on ImageNet-22k, our CvT-W24 obtains a top-1 accuracy of 87.7% on the ImageNet-1k val set. The positional encoding, a crucial component in existing Vision Transformers, can be safely removed in our model, simplifying the design for higher resolution vision tasks. ### Key features: - **CvT-13**: 224x224, 20M parameters, 4.5 GFLOPs, 81.6% top-1 accuracy. - **CvT-21**: 224x224, 32M parameters, 7.1 GFLOPs, 82.5% top-1 accuracy. - **CvT-13**: 384x384, 20M parameters, 16.3 GFLOPs, 83.0% top-1 accuracy. - **C

Overview: Poozle is an open-source single API for product integrations, providing a unified interface for integrating multiple tools like GitHub, Jira, Notion, Zendesk, and Salesforce. Key features: - Managed Authentication - Bi-directional data syncing - Automatic rate-limits & retries - Pagination - Unified data model - Webhooks (coming soon) - Built-in monitoring (coming soon) Tech stack: TypeScript Why it matters: Poozle simplifies the process of integrating various product tools, making it easier for companies to build new integrations without the need to manage multiple APIs and infrastructure.

Overview: DensePhrases is a text retrieval model that can return phrases, sentences, passages, or documents for natural language inputs. It leverages billions of dense phrase vectors from Wikipedia to efficiently search for answers to questions at the phrase level. Key features: - Phrase-level retrieval - Sentence-level retrieval - Passage-level retrieval - Document-level retrieval Tech stack: Python Why it matters: DensePhrases is a powerful tool for text retrieval tasks, enabling precise and efficient answers to natural language queries. Its ability to handle large-scale datasets makes it suitable for a wide range of applications, including knowledge-grounded dialogue and entity linking. Repository: princeton-nlp/DensePhrases Primary language: Python GitHub stars: 606 GitHub description: [ACL 2021] Learning Dense Representations of Phrases at Scale; EMNLP'2021: Phrase Retrieval Learns Passage Retrieval, Too https://arxiv.org/abs/2012.12624

Overview: MoCoGAN is a generative model designed to generate videos by decomposing motion and content, offering flexibility in generating videos with different actions and identities. Key features: - Separated representations of motion and content - Control over generated videos - Facial expressions generation from a facial expression database - Shape motion generation from a synthetically generated dataset - TaiChi videos generated from a large-scale TaiChi dataset Tech stack: Python Why it matters: MoCoGAN provides a novel approach to video generation by separating motion and content, offering versatility and flexibility in generating videos with different actions and identities, making it a valuable tool in the field of video synthesis and generation.

### Overview: Agentfy is a modular microservices architecture designed to process user requests and execute workflows across multiple social media platforms. It leverages LLMs to translate user intents into structured, goal-driven task chains, enabling intelligent workflow execution through multi-agent communication and collaboration. ### Key Features: - 🎯 **Buyer Targeting & Outreach**: Automate tasks across multiple social media platforms to find and engage with potential customers. - 📢 **Cross-Platform Promotions**: Send promotional content across multiple platforms to reach a wider audience. - 📝 **Content Transformation & Posting**: Transform ideas, videos, or trends into tailored posts, captions, and media across authorized accounts. - 🤖 **Automated Messaging & Support**: Respond to customer inquiries and support requests automatically. - 🧠 **Creator Discovery & Competitor Monitoring**: Identify and reach out to influencers and livestreamers for advertising and influencer marketing opportunities. ### Tech Stack: Python ### Why it Matters: Agentfy is designed to automate tasks across multiple social media platforms, making it an essential tool for digital marketers and social media managers looking to streamline their operations and increase efficiency.

Overview: Fabric is a creative coding and rapid prototyping environment that focuses on interactive visuals, image and video processing, and 3D content authoring. It is inspired by Apple's Quartz Composer ecosystem and aims to be a user-friendly tool for both beginners and advanced users. Key features: - Visual Node-based content authoring environment - SDK to load common interchange file formats - SDK to extend Fabric via a plugin architecture Tech stack: Swift, Metal, Xcode, Discord, and Satin. Why it matters: Fabric provides a powerful and intuitive environment for creative coding, rapid prototyping, and the creation of interactive visual experiences, making it a valuable tool for artists, developers, and anyone looking to explore the intersection of code and visual art.

### Overview: Sycamore is an AI-powered document processing engine designed for unstructured data, providing efficient partitioning, enrichment, and analysis capabilities. Leveraging Aryn's state-of-the-art DETR AI model, Sycamore excels in segmenting, labeling, and extracting data from complex documents, enhancing hybrid search and RAG applications. ### Key Features: - **Aryn DocParse Integration:** Seamlessly integrates with Aryn's state-of-the-art DETR AI model for high-quality segmentation and semantic preservation. - **DocSet Abstraction:** Scalable and robust abstraction for transforming and manipulating unstructured documents. - **Data Extraction and Enrichment:** Extracts tables, OCR, visual summaries, and performs performant Python data transforms. - **Vector Embeddings:** Quickly creates vector embeddings using your preferred AI model. - **Jupyter Notebook:** Helps in writing and iterating on jobs, and supports an OpenSearch hybrid search and RAG engine for testing. - **Ray Backend:** Scalable backend for efficient processing. ### Tech Stack: Python ### Why It Matters: Sycamore provides a powerful, AI-driven solution for processing and analyzing unstructured data, enhancing search and analytics capabilities, and improving the efficiency of hybrid search and RAG applications.

Overview: Busabase is an open-source database and workspace designed for AI agents, providing structured data, durable knowledge, reusable skills, runnable apps, and human review on the writes that matter. It is a local-first, self-hostable solution. Key features: - Structured data and knowledge management - Durable and shared workspace - Reusable skills and apps - Review and trust mechanisms for important writes - Agent statelessness and work durability Tech stack: TypeScript Why it matters: Busabase offers a centralized, shared environment for AI agents, ensuring that all agents have access to the same structured data, documents, skills, and apps, and that important writes are reviewed and trusted records. This helps maintain consistency and reliability across different agents, making the system more efficient and reliable. Repository: busabase/busabase GitHub stars: 602

Overview: Crafter is an open-world survival game designed to benchmark the capabilities of reinforcement learning agents. It features randomly generated 2D environments where players must forage, sleep, defend, collect materials, and build tools to survive. Key features: - Randomly generated 2D environments - Survival gameplay with foraging, sleeping, defending, collecting materials, and building tools - Evaluation of reinforcement learning agents on a wide range of capabilities Tech stack: Python Why it matters: Crafter provides a comprehensive benchmark for reinforcement learning agents, focusing on research challenges, meaningful evaluation, and efficient evaluation through iteration speed. Repository: danijar/crafter Primary language: Python GitHub stars: 604 GitHub description: Benchmarking the Spectrum of Agent Capabilities README excerpt: The repository contains a survival game environment for evaluating reinforcement learning agents. It features randomly generated 2D environments, survival gameplay, and evaluation metrics such as success rates and geometric mean score. The environment is compatible with the OpenAI Gym interface and can be installed using `pip3 install crafter`. For more information, refer to the research paper [Benchmarking the Spectrum of Agent Capabilities](https://arxiv.org/pdf/2109.06780.pdf).

Overview: ChatDB is an innovative project that aims to enhance Large Language Models (LLMs) by incorporating symbolic memory through a combination of an LLM and a set of SQL databases. This approach simulates complex multi-hop reasoning, aiming to improve the models' ability to perform intricate reasoning tasks. Key features: - Symbolic memory augmentation for LLMs - Integration of SQL databases - Continuous feature development - Validation on a synthetic dataset Tech stack: Python Why it matters: ChatDB offers a novel solution to enhance LLMs by leveraging symbolic memory, which is crucial for improving their ability to reason and perform complex tasks, making it a valuable addition to the field of AI and language models. Repository: huchenxucs/ChatDB Primary language: Python GitHub stars: 606 GitHub description: The official repository of "ChatDB: Augmenting LLMs with Databases as Their Symbolic Memory".

Overview: NextPlaid and ColGREP are multi-vector search engines designed to enhance code search capabilities, enabling developers to find relevant code snippets directly within their local environment. Key features: - **ColGREP**: A semantic code search tool that combines regex filtering with semantic ranking, focusing on local codebases. - **NextPlaid**: A multi-vector search engine that integrates with ColGREP, providing enhanced search capabilities for developers. Tech stack: Rust, leveraging Rust for building the search engine and its associated components. Why it matters: These tools streamline the process of finding relevant code snippets, making the development and debugging experience more efficient and accurate.

### Overview: VQAScore is a Python library designed to evaluate text-to-image, text-to-video, and text-to-3D models using a single line of code. It simplifies the process of benchmarking these models by providing a comprehensive evaluation framework. ### Key Features: - **Text-to-Image/Video/3D Model Evaluation**: Quickly assess the performance of text-to-image, text-to-video, and text-to-3D models. - **One-Line Scoring**: Utilizes Python to score models in a single command, making it highly efficient and user-friendly. - **Versatile Models**: Supports a wide range of models including CLIP-FlanT5, InstructBLIP, Qwen3.5, and others. ### Tech Stack: Python ### Why It Matters: VQAScore simplifies the evaluation process, making it accessible to researchers and developers without requiring extensive setup or knowledge of complex models. It allows for quick and accurate benchmarking, enabling more efficient model development and deployment.

Overview: YARP is a versatile library and toolkit designed for communication and device interfaces, used across various robotics projects, including humanoids and embedded systems. Key features: - YARP provides a robust communication framework for robotics applications. - It supports multiple platforms, including Linux, macOS, and Windows. - YARP includes tutorials for easy integration and usage. Tech stack: C++ Why it matters: YARP simplifies communication and device management in robotics, making it an essential tool for developers working on diverse robotics projects.

Overview: YT Navigator is an AI-powered application designed to help users efficiently navigate and search through YouTube channel content. It offers features such as searching through videos using natural language queries, chatting with channel content, and discovering relevant video segments with precise timestamps. Key features: - 🔍 Search through a channel's videos using semantic search - 💬 Chat with a channel's content - 📝 Extract insights from hours of content in seconds Tech stack: Python Why it matters: YT Navigator simplifies the process of extracting information from YouTube channels, making it an invaluable tool for researchers, students, content creators, and anyone seeking quick access to video content insights. Repository: wassim249/YT-Navigator Primary language: Python GitHub stars: 605 GitHub description: YT Navigator: AI-powered YouTube content explorer that lets you search and chat with channel videos using AI agents. Extract insights from hours of content in seconds with semantic search and precise timestamps.

### Overview: GraphRAG4OpenWebUI integrates Microsoft's GraphRAG technology into Open WebUI, providing advanced information retrieval capabilities for open web user interfaces. ### Key features: - **Local Search**: Utilizes GraphRAG technology for efficient local knowledge base retrieval. - **Global Search**: Searches information beyond local knowledge bases for comprehensive results. - **Tavily Search**: Integrates external Tavily search API for additional internet search capabilities. - **Full Model Search**: Combines all three search methods for the most comprehensive results. ### Tech stack: Python ### Why it matters: This project simplifies the integration of graph-based retrieval into open web environments, making it easier to build advanced information retrieval systems and search engines. Repository: win4r/GraphRAG4OpenWebUI GitHub stars: 606