Projects

### Overview: SocratiCode is an enterprise-grade codebase intelligence platform that provides zero-setup, local & private, hybrid semantic search, polyglot dependency graphs, symbol-level impact analysis, and interactive HTML viewer capabilities. It offers a 61% reduction in tokens, 84% fewer calls, and 37x faster performance compared to traditional methods. The platform is designed to be highly scalable and accessible, with a 3342 GitHub stars indicating its wide adoption and community support. ### Key features: - **Hybrid Semantic Search**: Enables precise and comprehensive search across multiple programming languages. - **Polyglot Dependency Graphs**: Provides a unified view of dependencies across different languages and frameworks. - **Symbol-Level Impact Analysis**: Identifies the impact of changes on specific code elements. - **Call-Flow Analysis**: Reveals the call relationships within the codebase, aiding in understanding the flow of execution. - **Interactive HTML Viewer**: Offers an intuitive interface for exploring and analyzing code. - **Cross-Project & Branch-Aware Search**: Seamlessly integrates across multiple projects and branches. - **DB/API/Infra Knowledge**: Provides comprehensive knowledge of database, API, and infrastructure details. ### Tech stack: The platform is built using TypeScript, leveraging modern JavaScript features and best practices. ### Why it matters: SocratiCode is essential for developers who need to quickly understand and analyze their entire codebase, including infrastructure, without the need for manual setup. Its advanced features and performance make it a valuable tool for improving code quality, reducing maintenance costs, and enhancing

Overview: This project hosts the code for implementing the Fully Convolutional One-Stage Object Detection (FCOS) algorithm, presented in the paper "FCOS: Fully Convolutional One-Stage Object Detection" at the 2019 International Conference on Computer Vision (ICCV). Key features: - FCOS completely avoids the complicated computation related to anchor boxes and all hyper-parameters of anchor boxes. - The very simple one-stage detector achieves much better performance (38.7 vs. 36.8 in AP with ResNet-50) than Faster R-CNN. - FCOS also requires less training hours (6.5h vs. 8.8h) than Faster R-CNN and takes 12ms less inference time per image than Faster R-CNN (44ms vs. 56ms). - Our best model based on ResNe(x)t-64x4d-101 and deformable convolutions achieves 49.0% in AP on COCO test-dev (with multi-scale testing). Tech stack: Python Why it matters: FCOS is a state-of-the-art one-stage detector that significantly improves upon Faster R-CNN, achieving better performance and faster training/testing times. It is particularly useful for real-time applications due to its low computational requirements and fast inference speed. Repository: tianzhi0549/FCOS Primary language: Python GitHub stars: 3347 GitHub description: FCOS: Fully Convolutional One-Stage Object Detection (

### Overview: FuXi is a fast, self-contained terminal AI coding agent that reads code, edits files, runs commands, and drives tools. It routes cost-awarely across LLM providers and supports automatic failover, making it a powerful tool for developers. ### Key features: - **Terminal-first**: FuXi operates entirely within the terminal, making it accessible on any platform. - **Provider-agnostic**: It supports multiple LLM providers, ensuring compatibility with various tools and services. - **Bring your own key**: Users can connect their own models or use FuXi-managed models. - **MCP client**: FuXi is built in Go and runs as a static binary, eliminating runtime dependencies. - **Self-updating**: The agent is continuously updated to ensure the latest features and improvements. ### Tech stack: FuXi is built using Go, making it fast and efficient. ### Why it matters: FuXi is designed to be a powerful and flexible tool for developers, providing a fast and efficient way to interact with code and tools, while also ensuring data privacy and security.

### Overview: This repository contains a collection of practice and tutorial-style notebooks designed to help developers practice and learn machine learning techniques using Python. The notebooks cover a wide variety of topics and are written in Jupyter Notebook format, making them ideal for hands-on learning. ### Key Features: - Detailed explanations and examples for NumPy, Pandas, Matplotlib, Seaborn, and other essential libraries. - Comprehensive coverage of various machine learning techniques, including but not limited to linear regression, decision trees, neural networks, and more. - Practical examples and exercises to help developers understand and apply machine learning concepts. - Written in a tutorial-style format, making it easy to follow and learn from. - Suitable for developers looking to improve their skills in machine learning and data science. ### Tech Stack: - Python 3.6+ - NumPy - Pandas - Matplotlib - Seaborn - SciPy - Statsmodels - Flask - WTForms - TensorFlow - Keras - pdpipe ### Why It Matters: These notebooks provide a practical and accessible way for developers to learn and apply machine learning techniques using Python. They are designed to be used as a reference or as a starting point for more in-depth study. The tutorials are written in a clear and concise manner, making them ideal for both beginners and experienced developers looking to improve their skills. ### Repository: tirthajyoti/Machine-Learning-with-Python - Primary language: Jupyter Notebook - GitHub stars: 3338 - GitHub description: Practice and tutorial

### Overview: This comprehensive survey and collection of resources on Context Engineering spans from prompt engineering to production-grade AI systems, focusing on LLMs and AI agents. It includes hundreds of papers, frameworks, and implementation guides, covering topics such as LLMs, AI agents, agent runtimes, memory systems, protocols, coding agents, and observability stacks. ### Key features: - **Context Engineering**: The evolution from static prompting to dynamic, context-aware AI systems. - **Agent Era**: Shift from "how to pack the best prompt" to how agent systems manage runtime state, memory, tools, protocols, approvals, and long-horizon execution. - **Agent Harnesses and Runtime Systems**: Planning, subagents, checkpoints, sandboxes, and human approval loops. - **Context Management in Production**: Compaction, caching, artifact-backed context, and scoped instruction loading. - **Memory Artifacts and Portability**: Persistent memory, memory interchange formats, persona packaging, and project memory. ### Tech stack: - Comprehensive survey and collection of resources on Context Engineering. ### Why it matters: This repository serves as a valuable resource for developers and researchers interested in the evolution of AI systems, focusing on the practical aspects of context engineering and agent management.

Overview: This repository provides hands-on CUDA kernels, CUTLASS/CuTe notes, Triton examples, PTX ISA notes, PyTorch internals notes, and LLM inference/training optimization material. It is a collection of my study and engineering notes on GPU systems work. Key features: - CUDA kernels for reduce, softmax, elementwise, GEMV, indexing, atomic add, upsampling, and linear attention. - Notes and code from the CUDA-MODE lecture series. - CUTLASS and CuTe DSL notes, including GEMM, TMA, WGMMA, swizzling, and instruction-level material. - Triton kernels, PyTorch interop examples, and meetup notes. - LLM serving, training, and systems optimization notes. - PyTorch internals and CUDA-related notes. - PTX ISA study notes. - Small helper scripts. - Older material kept for reference. Tech stack: Cuda Why it matters: This repository offers practical resources for optimizing CUDA kernels, improving AI infrastructure, and enhancing deep learning frameworks, making it valuable for developers working on GPU systems and AI optimization.

### Overview: Mini-Gemini is a framework designed to mine the potential of multi-modality vision language models, supporting a series of dense and MoE Large Language Models (LLMs) from 2B to 34B with image understanding, reasoning, and generation simultaneously. ### Key features: - **Model Support**: Supports LLaVA-based models from 2B to 34B. - **Demo**: Available online at <http://103.170.5.190:7860/>. - **Code and Models**: Released on GitHub at <https://github.com/dvlab-research/MGM>. - **Data**: Available on Hugging Face at <https://huggingface.co/collections/YanweiLi/mgm-data-660463ea895a01d8f367624e>. - **Paper**: Published on arXiv at <https://arxiv.org/pdf/2403.18814.pdf>. ### Tech stack: The framework is built using Python and relies on transformers, fastai, and other libraries. ### Why it matters: Mini-Gemini provides a comprehensive solution for multi-modality language models, enabling more accurate and efficient interactions between text and images.

### Overview: Eiten is an open-source toolkit designed to implement various statistical and algorithmic investing strategies, including Eigen Portfolios, Minimum Variance Portfolios, Maximum Sharpe Ratio Portfolios, and Genetic Algorithms-based Portfolios. It allows users to build their own portfolios using their own set of stocks, ensuring they can outperform the market. ### Key Features: - **Eigen Portfolios**: A portfolio optimization technique that aims to maximize the portfolio's expected return for a given level of risk. - **Minimum Variance Portfolios**: A portfolio optimization method that seeks to minimize the portfolio's risk by diversifying across multiple assets. - **Maximum Sharpe Ratio Portfolios**: A portfolio optimization strategy that aims to maximize the ratio of the expected return to the risk, providing a balance between risk and return. - **Genetic Algorithms-based Portfolios**: A portfolio optimization approach that uses evolutionary algorithms to find the optimal portfolio weights. ### Tech Stack: Python ### Why It Matters: Eiten provides a rigorous testing framework that ensures the confidence in the performance of the portfolios built by the user. The toolkit allows anyone to build their own portfolios using their own set of stocks, making it accessible to a wide range of investors. ### Repository: tradytics/eiten ### GitHub Stars: 3322 ### Primary Language: Python ### GitHub Description: Statistical and Algorithmic Investing Strategies for Everyone ### Files Description: - **eiten**: Main folder. - **&boxur;&nbsp; figures**: Figures for this GitHub repository. - **&boxur;&nbsp; stocks**:

### Overview: This leaderboard evaluates the performance of language models in producing hallucinations when summarizing short documents, using Vectara's Hallucination Evaluation Model (HHEM). ### Key features: - **Model Comparison**: Tracks performance across various LLMs. - **Hallucination Rate**: Measures the frequency of hallucinations introduced by each model. - **Factual Consistency Rate**: Ensures the summaries are factually accurate. - **Answer Rate**: Evaluates the accuracy of the summaries. - **Average Summary Length**: Indicates the length of the summaries produced. ### Tech stack: Python ### Why it matters: This leaderboard helps developers and researchers understand the strengths and weaknesses of different LLMs in generating accurate and factually consistent summaries, thereby guiding improvements and optimizations in model design and training.

Overview: ROOT is a powerful software package designed for the analysis, storage, and visualization of scientific data, particularly in the field of particle physics. It is optimized for handling large datasets and provides a wide range of functionalities including histogramming, curve fitting, statistical modeling, and parallel processing. Key features: - Scientific data storage and analysis - Columnar ROOT format for efficient data handling - Arbitrary dimension histogramming capabilities - Curve fitting and statistical modeling - Parallel processing framework RDataFrame - Performance-critical C++ implementation - Cling, a C++ compliant interpreter for rapid prototyping Tech stack: C++ Why it matters: ROOT is essential for scientific data analysis and visualization, enabling researchers to efficiently store, process, and visualize large datasets, thereby maximizing the impact of their work. Its performance and reliability make it a crucial tool in the field of particle physics, where data analysis is critical for understanding fundamental particles and interactions. Repository: root-project/root GitHub stars: 3318 GitHub description: The official repository for ROOT: analyzing, storing and visualizing big data, scientifically.

### Overview: Sidekick is a macOS app that allows users to chat with a local LLM, responding with information from files, folders, and websites on their Mac. It uses llama.cpp and supports modern GGUF local models. ### Key Features: - **Local First**: Sidekick is a local-first application, using built-in inference engines for local models and accommodating OpenAI compatible APIs for additional model options. - **Resource Access**: Users can access files, folders, and websites from their experts, which can be individually configured to contain resources related to specific areas of interest. - **RAG Integration**: Sidekick uses RAG (Retrieval Augmented Generation) to find relevant information, allowing unlimited resources to be accessed. - **Web Search**: Sidekick can respond with the latest information using web search, speeding up research. - **Bring Your Own API Key**: Users can bring their own key for OpenAI compatible APIs, allowing access to additional remote models. ### Tech Stack: Swift ### Why it Matters: Sidekick is a powerful tool for users who need a local-first approach to AI, allowing them to access information directly from their Mac without the need for any additional software. It is particularly useful for students and researchers who need to quickly find and analyze information related to their work. ### Repository: johnbean393/Sidekick ### GitHub Stars: 3319 ### GitHub Description: A native macOS app that allows users to chat with a local LLM that can respond with information from files, folders and websites on your Mac without installing any other software.

Overview: This repository is a curated list of open source GitHub repositories related to ChatGPT, the OpenAI API, and Codex. It provides a searchable tool for finding and filtering repositories based on specific criteria such as topic, category, language, and more. Key features: - Searchable repositories directly from Claude Code or Codex - RAG frameworks, agent libraries, chatbot UIs, CLI tools, and more - Filter by category or language - Multilingual queries - Scored results by relevance, star count, and project quality Tech stack: Python Why it matters: This repository serves as a valuable resource for developers looking to explore and utilize open-source ChatGPT and Codex-related projects. The curated list and search functionality make it easy to find and evaluate the latest and most relevant tools and repositories in the field.

Overview: Reflexion: Language Agents with Verbal Reinforcement Learning is a project that presents a novel approach to language agents using verbal reinforcement learning, as detailed in the NeurIPS 2023 paper by Noah Shinn et al. Key features: - Code, demos, and log files for the NeurIPS 2023 paper - LeetcodeHardGym repository for additional resources - HotPotQA notebooks for reasoning experiments - AlfWorld repository for decision-making experiments - Available agent types: ReAct, CoT_context, CoT_no_context, and Reflexion strategies: NONE, LAST_ATTEMPT, REFLEXION, LAST_ATTEMPT_AND_REFLEXION - Available run parameters: num_trials, num_envs, run_name, use_memory, is_resume, resume_dir, start_trial_num - Logs are stored in `./root/<run_name>` for reasoning, `./root/<run_name>` for decision-making, and `./root/<run_name>` for programming Tech stack: Python Why it matters: This project showcases advancements in language agents using verbal reinforcement learning, providing a valuable resource for researchers and developers interested in AI and reinforcement learning. Repository: noahshinn/reflexion

Overview: 本项目是一个基于LangChain和ChatGLM-6B系列模型的本地知识库自动问答系统，支持上传多种文本格式文件，包括txt、docx、md和pdf等，提供包括ChatGLM-6B系列、Belle系列等模型文件以及GanymedeNil/text2vec-large-chinese、nghuyong/ernie-3.0-base-zh、nghuyong/ernie-3.0-nano-zh等Embedding模型。 Key features: - 支持上传多种文本格式文件 - 提供本地知识库应用 - 模型包括ChatGLM-6B系列、Belle系列等 - 提供GanymedeNil/text2vec-large-chinese、nghuyong/ernie-3.0-base-zh、nghuyong/ernie-3.0-nano-zh等Embedding模型 Tech stack: Python Why it matters: 本项目利用LangChain和ChatGLM-6B系列模型，提供了一个强大的本地知识库应用，适用于各种文本格式文件的自动问答，同时支持多种Embedding模型，使得应用更加灵活和高效。

### Overview: This repository is a curated list of AI residency programs, offering opportunities for individuals to engage in research and development within the field of artificial intelligence. ### Key features: - **All Year**: Programs available throughout the year, including internships and bootcamps. - **2025**: Programs such as AIML Resident, Anthropic AI Safety Fellowship, and AI Residency @ Cloudwalk. - **2024**: Programs like Apple AIML Resident, Fatima Fellowship, and SandboxAQ Residency Program. - **2023**: Programs such as Microsoft Research: Postdoc Residency Programme, Cambridge Residency Programme, and Meta AI Residency. - **2022**: Programs including OpenAI Residency, Hugging Face, NVIDIA AI Research Residency, Novartis AI for Life Residency Program, and Allen AI Predoctoral Young Investigator. ### Tech stack: The repository uses Markdown for its content and GitHub for hosting. ### Why it matters: These residency programs provide hands-on experience in AI research, fostering innovation and contributing to the field's advancement.

Overview: MLJAR is a Python package designed for Automated Machine Learning (AutoML) on tabular data. It includes feature engineering, hyper-parameter tuning, explanations, and automatic documentation, making it accessible for developers. Key features: - Feature Engineering - Hyper-Parameter Tuning - Model Explanations - Automatic Documentation Tech stack: Python Why it matters: MLJAR simplifies the process of building and deploying machine learning models, making it easier for developers to automate the machine learning pipeline without extensive knowledge of machine learning algorithms and frameworks. Repository: mljar/mljar-supervised Primary language: Python GitHub stars: 3303 GitHub description: Python package for AutoML on Tabular Data with Feature Engineering, Hyper-Parameters Tuning, Explanations and Automatic Documentation README excerpt: # MLJAR Automated Machine Learning for Humans [](https://github.com/mljar/mljar-supervised/actions/workflows/run-tests.yml) [](https://badge.fury.io/py/mljar-supervised) [](https://anaconda.org/conda-forge/mljar-supervised) [![Anaconda-Server Badge](https://anaconda.org/conda-forge/mljar-supervised/badges/license

Overview: `claude-tap` is a local proxy and trace viewer designed to intercept and inspect the API traffic from various AI coding agents, including Claude Code, Codex CLI, Gemini CLI, Cursor CLI, OpenCode, Kimi/Kimi Code, Pi, and Hermes. It allows developers to view and compare the real API traffic and agent context, such as system prompts, conversation history, tool schemas, tool calls, streaming responses, token usage, and request diffs. Key features: - Intercept and inspect API traffic from multiple AI coding agents. - View and compare real agent runs with detailed context and evidence. - Generate a portable artifact for each run, capturing the local trace session. Tech stack: Python Why it matters: `claude-tap` provides a powerful tool for developers to inspect and debug the context and behavior of AI coding agents, ensuring accurate and detailed analysis of the interactions between prompts, messages, tool calls, and more.

### Overview: RAG Web UI is an intelligent dialogue system based on RAG technology, designed to help build knowledge-based Q&A systems. It supports multiple LLM deployment options and provides OpenAPI interfaces for easy knowledge base access. ### Key Features: - **Intelligent Document Management**: Supports multiple document formats (PDF, DOCX, Markdown, Text) and automatically processes them for vectorization. - **Advanced Dialogue Engine**: Precise retrieval and generation based on RAG, supporting multi-turn contextual dialogue and reference citations. - **Robust Architecture**: Frontend-backend separation, distributed file storage, and high-performance vector database (ChromaDB, Qdrant). ### Tech Stack: TypeScript ### Why It Matters: RAG Web UI leverages RAG technology to create intelligent Q&A systems, making it a valuable tool for knowledge-based applications. Its modular design and flexibility make it suitable for various deployment scenarios, including cloud services and local model deployment. Repository: rag-web-ui/rag-web-ui GitHub Stars: 3301 GitHub Description: RAG Web UI is an intelligent dialogue system based on RAG (Retrieval-Augmented Generation) technology.

Overview: Zeron is a native control plane designed for various coding agents, allowing users to manage and synchronize their sessions locally or across multiple devices. Key features: - Local control and multi-device synchronization - Desktop app for macOS, Windows, and Linux - CLI (Command Line Interface) for server environments - Multi-device synchronization for trusted users - Sponsors and sponsorship options Tech stack: Rust Why it matters: Zeron simplifies the management and synchronization of coding agents, making it easier for developers to control and collaborate with their agents efficiently. Repository: zeronsh/zeron GitHub stars: 3285

### Overview: Haiku is a JAX-based neural network library designed to simplify the process of building and managing neural networks, making it easier to integrate with existing JAX codebase. ### Key Features: - **Object-Oriented Programming Models**: Haiku provides a simple yet powerful way to build neural networks using familiar object-oriented programming models. - **Pure Function Transformations**: It transforms functions that use object-oriented, functionally "impure" modules into pure functions, allowing them to be used with JAX's JIT, grad, and pmap transformations. - **Module Abstraction (`hk.Module`)**: Haiku's `hk.Module` is a Python object that holds references to parameters, other modules, and functions that apply functions on user inputs. - **Function Transformation (`hk.transform`)**: `hk.transform` turns functions that use these object-oriented, functionally "impure" modules into pure functions, enabling them to be used with JAX's transformation capabilities. ### Tech Stack: Haiku is built on top of JAX, a numerical computing library that combines NumPy, automatic differentiation, and first-class GPU/TPU support. ### Why It Matters: Haiku simplifies the process of building and managing neural networks, making it easier to integrate with existing JAX codebase. It is designed to be a library, not a framework, and does not reinvent the wheel, focusing on specific simplifications while leveraging the existing JAX ecosystem.