Coding & Development

Mergoo is an open-source Python library designed to simplify the process of merging multiple Large Language Model (LLM) experts and then efficiently training the resulting merged LLM. It enables users to integrate knowledge from different generic or domain-specific LLM experts, supporting methods such as Mixture-of-Experts (MoE) and Mixture-of-Adapters (MoA). The library offers flexible merging for each layer and supports popular base models like Llama (including LLaMa3), Mistral, Phi3, and BERT. It is compatible with various trainers including Hugging Face Trainer, SFTrainer, and PEFT, and can run on CPU, MPS, and GPU devices. Mergoo allows for training choices ranging from only the Router of MoE layers to fully fine-tuning the merged LLM.

DIG (Dive into Graphs) is a comprehensive open-source library designed for graph deep learning research. Unlike basic graph deep learning libraries, DIG offers a unified testbed for advanced, research-oriented tasks such as graph generation, self-supervised learning on graphs, explainability of Graph Neural Networks, deep learning on 3D graphs, and graph out-of-distribution. It provides unified implementations of data interfaces, common algorithms, and evaluation metrics, allowing researchers to easily implement their own methods and compare them against baseline methods using common datasets and metrics without extensive effort. The library supports various research directions including Graph Augmentation and Fair Graph Learning, and is built on PyTorch Geometric (PyG).

DeepLearningPython is a GitHub repository that offers updated scripts from neuralnetworksanddeeplearning.com, specifically tailored for Python 3.5.2 and integrated with the Theano deep learning library, including CUDA support. This resource provides a practical foundation for individuals looking to learn and implement neural networks. The repository includes three distinct network implementations (network.py, network2.py, network3.py) from the original book, all runnable via a single testing file, test.py. This setup allows users to easily train and evaluate different network configurations, with examples and comments linking back to specific chapters of the book. It's an excellent tool for hands-on learning and experimentation in deep learning.

IOVISION is an offshore engineering and software development company based in Tunisia, specializing in artificial intelligence and software development services. They offer customized AI-based solutions, web and mobile application development, big data analysis, and software development outsourcing. Their services aim to help businesses integrate AI into operations, digitize processes, and make informed decisions through data analysis. IOVISION also develops its own products, including Finispia (FinTech), MooMe (AI for dairy farm management), and Stile (AI-based personal shopping app). They serve various industries such as Media & Entertainment, Transportation & Logistics, Finance, Agriculture, Healthcare, and Retail/E-commerce, providing multidisciplinary expertise for product development and consulting.

fin-ml is an open-source GitHub repository dedicated to providing machine learning and data science blueprints specifically tailored for finance. It features a collection of Jupyter notebooks that contain code for case studies from the O'Reilly book "Machine Learning and Data Science Blueprints for Finance." Users can clone the repository to run these notebooks locally, allowing for hands-on experimentation with the code. The repository covers a wide range of financial applications, including trading strategies, portfolio management, derivatives pricing, asset price prediction, fraud detection, loan default probability prediction, and chatbot development. It also categorizes notebooks by machine learning types such as supervised learning, unsupervised learning, reinforcement learning, and natural language processing, making it a comprehensive resource for developers and data scientists in the fintech domain.

GRPO-Zero is an open-source project that provides a from-scratch implementation of DeepSeek R1's Group Relative Policy Optimization (GRPO) algorithm. This tool is specifically designed for training large language models using reinforcement learning, emphasizing minimal dependencies and efficient GPU memory usage. It supports training on GPUs with limited VRAM, such as a 24GB RTX 4090, by offloading the optimizer to the CPU, incurring only a small overhead. Key improvements over the original GRPO include token-level policy gradient loss, removal of KL Divergence for reduced GPU memory, and optional overlong episode filtering to stabilize training. The project demonstrates its capabilities by training Qwen2.5 models on a CountDown task, where the model learns to generate mathematical expressions and reasoning.

MSTR specializes in providing direction and results with AI, translating complex business challenges into actionable advice and custom software solutions. From initial ideas to full-scale AI application deployment, MSTR assists organizations in working smarter, faster, and more efficiently. Their services include AI Consultancy for strategic guidance and training, Product Development for bespoke AI software integrated into existing processes, and SaaS Development for complete, domain-specific AI products. MSTR's team of AI consultants, data scientists, and engineers possesses the expertise to address diverse AI requirements, helping businesses step into a new era of operations.

OpenxAI Network is a pioneering decentralized AI ecosystem designed to make intelligence open, programmable, and liquid. It functions as the first P2P permissionless AI protocol, allowing users to create, own, consume, and exchange AI as a shared global utility. The platform aims to disrupt centralized AI by providing an end-to-end permissionless AI stack, from hardware to on-chain AI. Key components include X Engine for decentralized compute, X Studio for web-based deployment and management, a Playground for instant AI service testing, and a Marketplace for tokenized models and datasets. OpenxAI emphasizes censorship resistance, on-chain monetization, and community governance, offering a robust environment for developers to build and commercialize AI applications without traditional gatekeepers.

HiddenLayer is a lightweight Python library designed to visualize neural network graphs and training metrics for popular deep learning frameworks like PyTorch, TensorFlow, and Keras. It's particularly useful for developers and data scientists working in Jupyter Notebook environments, offering a simpler alternative to more advanced tools like TensorBoard for quick analysis. Key features include rendering high-level network architectures, customizable graph expressions for hiding and folding nodes, and plotting training metrics such as loss, accuracy, and weight histograms. It can also be used outside Jupyter Notebook to save graph snapshots or display metrics in a separate window, making it a versatile tool for understanding and debugging neural networks.

Neuronika is a machine learning framework built entirely in Rust, emphasizing ease of use, rapid prototyping, and performance. At its core, Neuronika utilizes reverse-mode automatic differentiation, enabling the creation of dynamically changing neural networks with minimal effort and overhead through a lean, imperative, and define-by-run API. The framework leverages the power of the Rust language to offer an intuitive and efficient interface without the need for Foreign Function Interfaces (FFI). It supports GPU-accelerated primitives via CUDA, serialization with Serde, and transparent BLAS support for optimized matrix multiplication. Neuronika is currently in active development, with breaking changes expected as it evolves.

json_repair is a Python library designed to automatically fix malformed JSON data, a common issue when working with outputs from Large Language Models (LLMs), APIs, logs, or user input. It addresses a wide range of syntax errors, including missing quotes, commas, brackets, unescaped characters, and truncated values. The tool can serve as a direct replacement for Python's built-in `json.loads()` and `json.load()` functions, offering a robust fallback mechanism. It also supports schema-guided repairs, allowing users to define a JSON Schema or Pydantic v2 model to ensure outputs conform to a specific structure, filling missing values or coercing types. The library is open-source and available on GitHub, providing a flexible solution for developers needing reliable JSON parsing.

ReDX Technologies positions itself as a reliable partner for digital transformation, specializing in High-Performance Computing (HPC) and Artificial Intelligence (AI) solutions. The company aims to support organizations in leveraging these advanced technologies. Their offerings include HPC & AI services, HPC & AI products, and HPC/AI procurement projects, indicating a comprehensive approach to integrating these technologies into businesses. While specific features are not detailed, the focus is on providing support and solutions for digital initiatives, suggesting a consultancy and implementation role rather than a direct software product.

learn-claude-code is a GitHub repository that serves as a masterclass in harness engineering for AI agents. It focuses on the principle that agency comes from the model's training, not from external code orchestration, and teaches users how to build the 'vehicle' (harness) for these intelligent 'drivers' (models). The repository breaks down the architecture of Claude Code into 12 progressive sessions, covering mechanisms like agent loops, tools (bash, read, write, edit, browser), on-demand skill loading, context compression, subagent spawning, and task systems. It emphasizes building the environment for AI intelligence to operate effectively, rather than trying to program intelligence itself. This resource is ideal for developers looking to understand and implement robust agent harnesses for various domains, from coding to agriculture and healthcare.

ML-Recipes is a comprehensive, open-source collection of stand-alone Python examples for various machine learning algorithms. It serves as a practical resource for developers and data scientists looking to understand and implement ML concepts. The repository includes recipes for Multi-armed bandit (MAB) algorithms like Epsilon greedy and Thompson sampling, Artificial Neural Networks (ANN) such as Adaptive Resonance Theory and Multi-Layer Perceptron, Markov Decision Process (MDP) with Value Iteration, and Dimensionality Reduction (DR) techniques including Principal Component Analysis and Eigenface. Each recipe is designed to be reasonably small and includes usage instructions and results, making it easy to integrate into projects or use for learning purposes.

Akinciborg Security provides comprehensive penetration testing and ethical hacking services, specializing in web applications, APIs, and cloud environments. The service aims to secure digital assets by proactively identifying vulnerabilities such as SQL Injection, XSS, RCE, and business logic flaws. Akinciborg offers detailed security assessments, including OWASP Top 10 testing, API authentication and rate limiting verification, and in-depth frontend and backend security analysis. Additionally, they provide secure code review to detect flaws missed by automated tools, covering source code analysis, cryptography implementation, and third-party library audits. With transparent pricing packages and a rigorous methodology, Akinciborg helps organizations of all sizes enhance their security posture.

MCP-Framework is an open-source TypeScript framework designed for building Model Context Protocol (MCP) servers. It provides a robust architecture with out-of-the-box features like automatic directory-based discovery for tools, resources, and prompts. Developers can leverage powerful MCP abstractions to define these components elegantly, benefiting from TypeScript-first development with full type safety. The framework supports multiple transport options including stdio, SSE, and HTTP Stream, and includes built-in authentication for SSE endpoints using OAuth 2.1, JWT, or API Keys. A command-line interface (CLI) simplifies project creation, adding tools, prompts, and resources, and includes comprehensive validation for tool schemas to ensure proper documentation and functionality.

Stottler Henke Associates, Inc. has specialized in artificial intelligence since 1988, providing advanced software systems to tackle problems that traditional approaches cannot solve. Their expertise spans several key areas, including Education & Training, Planning & Scheduling, Decision Support, Knowledge Management & Discovery, and Autonomous Systems. The company offers a comprehensive suite of services, from technology consulting and feasibility studies to rapid prototyping, software development, implementation, and technology transfer, ensuring successful deployment of effective, long-term AI solutions. Notable products include Aurora for dynamic scheduling, InfoTracker for knowledge management, and DataMontage for data visualization. Stottler Henke is an employee-owned company, serving both corporations and government agencies like Boeing and NASA.

neurojs is an open-source JavaScript framework designed for deep learning and reinforcement learning applications within the browser environment. While it mainly focuses on reinforcement learning, it is versatile enough for various neural network-based tasks. The library includes practical examples and demos, such as a 2D self-driving car visualization, to showcase its capabilities. It supports advanced features like uniform and prioritized replay buffers, advantage-learning, and models such as deep-q-networks and actor-critic (via deep-deterministic-policy-gradients). neurojs also allows for binary import and export of network configurations, including weights, and is built for high performance. However, development on neurojs is no longer actively maintained, with the recommendation to use more general frameworks like TensorFlow-JS.

O-CNN is an open-source project from Microsoft that provides an implementation of Octree-based Convolutional Neural Networks (O-CNN) for various 3D shape analysis tasks. This framework is designed for researchers and developers working with 3D data, offering solutions for shape classification, retrieval, segmentation, autoencoding, and completion. It also supports unsupervised pretraining and Image2Shape functionalities. The repository includes code for different deep learning frameworks like PyTorch, TensorFlow, and Caffe, with recent updates focusing on PyTorch-based O-CNNs, including UNet and ResNet architectures. O-CNN has demonstrated strong performance on benchmarks like ScanNet and ModelNet40, often surpassing other state-of-the-art approaches.

RemoteCLIP is the official repository for the paper "RemoteCLIP: A Vision Language Foundation Model for Remote Sensing." This tool addresses limitations in existing remote sensing models by learning robust visual features with rich semantics and aligned text embeddings, crucial for retrieval and zero-shot applications. It leverages data scaling and conversion of heterogeneous annotations, incorporating UAV imagery to create a significantly larger pre-training dataset. RemoteCLIP supports diverse downstream tasks including zero-shot image classification, linear probing, k-NN classification, few-shot classification, image-text retrieval, and object counting, consistently outperforming baseline foundation models across various scales and datasets.

Practical-Deep-Learning-for-Coders-2.0 offers a comprehensive collection of notebooks designed for the "A walk with fastai2" Study Group and Lecture Series. This resource is ideal for individuals looking to delve into practical deep learning, covering key areas such as computer vision, tabular neural networks, and natural language processing. The course, which includes live-streamed lectures and project work, provides a structured learning path for undergraduates and others interested in the fastai framework. While the notebooks are now hosted on a new GitHub repository, this original repository serves as a valuable archive of the course material, offering insights into various deep learning applications and techniques.

ruby-fann is a Ruby Gem designed to interface with the FANN (Fast Artificial Neural Network) library, allowing Ruby and Rails developers to integrate neural network capabilities into their applications. This open-source library supports the implementation of both fully-connected and sparsely-connected artificial neural networks. It is lauded for its ease of use, versatility, and speed, with most of the heavy lifting performed natively. The gem provides functionalities for training neural networks with custom data, saving and loading trained networks, and implementing custom training procedures via callback methods, making it a robust solution for AI application development in Ruby environments.

Transformer Engine (TE) is an open-source library developed by NVIDIA for significantly accelerating Transformer models on NVIDIA GPUs. It achieves this by leveraging 8-bit and 4-bit floating point (FP8 and FP4) precision on Hopper, Ada, and Blackwell GPUs, including MXFP8 and NVFP4 formats on Blackwell. This results in improved performance and reduced memory utilization during both training and inference processes. TE provides highly optimized building blocks for popular Transformer architectures and an automatic mixed precision-like API that integrates seamlessly with existing framework-specific code. It also offers a framework-agnostic C++ API for broader integration, simplifying mixed-precision training for users by internally managing scaling factors.

TreeQuest is an open-source Python library designed for advanced tree search algorithms, particularly useful for scaling Large Language Model (LLM) inference. It offers a flexible API that allows for customizable node generation and scoring logic, making it adaptable to various applications. The library implements AB-MCTS-A (ABMCTS with Node Aggregation) and AB-MCTS-M (ABMCTS with Mixed Models) algorithms, as well as Multi-LLM AB-MCTS support. Key features include checkpointing and resuming searches, an ask-tell interface for batched sampling, and visualization utilities to render search trees. TreeQuest is ideal for developers and researchers working on optimizing LLM performance and exploring complex decision-making processes.