Coding & Development

SiT (Scalable Interpolant Transformers) offers an official PyTorch implementation for exploring advanced generative models. Built on the foundation of Diffusion Transformers (DiT), SiT introduces an interpolant framework that allows for flexible connections between distributions, surpassing DiT's performance on the conditional ImageNet 256x256 benchmark with identical backbones and parameters. This repository includes pre-trained class-conditional SiT models, a training script utilizing PyTorch DDP, and sampling code with various configurable options for ODE and SDE samplers. Researchers and developers can leverage SiT to experiment with discrete vs. continuous time learning, different model predictions, interpolant choices, and deterministic or stochastic sampling strategies.

Scikit-LLM provides a seamless integration of powerful large language models (LLMs) such as ChatGPT into the scikit-learn ecosystem, enabling enhanced text analysis tasks. This tool is designed for data scientists and machine learning engineers who wish to leverage advanced natural language processing capabilities directly within their familiar scikit-learn workflows. It simplifies the process of incorporating LLMs for tasks like zero-shot text classification, as demonstrated by its quick start example. Scikit-LLM is an open-source project available on GitHub, fostering community contributions and support. It aims to bridge the gap between traditional machine learning frameworks and the latest advancements in large language models, making sophisticated NLP more accessible for practical applications.

Soprano is an ultra-lightweight, on-device text-to-speech (TTS) model designed for expressive, high-fidelity speech synthesis at unprecedented speed. It boasts features like up to 20x real-time generation on CPU and 2000x real-time on GPU, lossless streaming with low latency, and minimal memory usage with a compact 80M parameter architecture. Soprano supports infinite generation length with automatic text splitting and crystal clear audio generation at 32kHz. It offers widespread support for CUDA, CPU, and MPS devices on Windows, Linux, and Mac, and provides an OpenAI-compatible endpoint, ONNX, WebUI, CLI, and Python script for easy and production-ready inference.

Suno AI Bark is an open-source, transformer-based text-to-audio model developed by Suno. It excels at generating highly realistic, multilingual speech, as well as other audio elements like music, background noise, and simple sound effects. Unlike conventional text-to-speech models, Bark is fully generative and can produce nonverbal communications such as laughing, sighing, and crying. It supports over 100 speaker presets across various languages and can automatically determine language from input text, even attempting native accents for code-switched text. The model is available for commercial use and can be integrated via Python or the Hugging Face Transformers library, offering flexibility for developers and researchers.

Text Generation Inference (TGI) is an open-source toolkit designed for deploying and serving Large Language Models (LLMs) with high performance. Developed by Hugging Face, it's used in production for services like Hugging Chat and the Inference API. TGI supports popular open-source LLMs including Llama, Falcon, and BLOOM, offering features such as tensor parallelism for faster inference on multiple GPUs, token streaming, and continuous batching for increased throughput. It also includes optimized transformers code with Flash Attention and Paged Attention, various quantization methods (bitsandbytes, GPT-Q, AWQ, Marlin, fp8), and hardware support for Nvidia, AMD, Inferentia, Intel GPU, Gaudi, and Google TPU. While TGI is now in maintenance mode, it has influenced the development of other optimized inference engines like vLLM and SGLang, which Hugging Face now recommends.

stable_diffusion.openvino is an open-source implementation of text-to-image generation using Stable Diffusion, specifically designed for efficient performance on Intel CPUs or GPUs. This tool allows users to generate images from text descriptions, offering capabilities like text-to-image, image-to-image, and inpainting. It supports various parameters for fine-tuning image generation, including model selection, inference device, random seed, guidance scale, and initial image strength. The project provides clear instructions for installation on Linux, Windows, and MacOS, requiring Python <= 3.9.0 and OpenVINO™ Development Tools. Performance benchmarks are included, showcasing its efficiency across different Intel processors.

KServe is a standardized, distributed platform designed for deploying both generative and predictive AI inference models on Kubernetes. It offers a unified solution for managing AI workloads, from quick deployments to enterprise-scale applications. Key features for generative AI include optimized backends for LLMs (vLLM, llm-d), OpenAI-compatible inference protocols, GPU acceleration, intelligent model caching, KV cache offloading, and request-based autoscaling. For predictive AI, KServe supports multiple frameworks like TensorFlow, PyTorch, scikit-learn, and XGBoost, along with intelligent routing, advanced deployments like canary rollouts, and model explainability. It also provides advanced monitoring capabilities and cost efficiency through scale-to-zero functionality.

Add Innovations Pvt Ltd is an NCR-based technology company specializing in AI-based vision systems and vision consulting. They offer a range of solutions including machine vision systems, scientific research optics, opto-imaging zoom lens systems, and AI-based image processing software. Their services focus on transforming processes through computer vision and deep learning, enabling error-free inspections, precision measurement, surface inspection, and vision-controlled robotics. Add Innovations aims to provide affordable machine vision solutions for complex technologies, with applications spanning appliance label inspection, spring measurement, automotive assembly inspection, and manufacturing.

Larq is an open-source deep learning library specifically designed for training neural networks with extremely low precision weights and activations, such as Binarized Neural Networks (BNNs). Traditional deep neural networks often use higher precision (32, 16, or 8 bits), making them large, slow, and power-hungry, which limits their application in resource-constrained environments. Larq addresses this by providing a framework to build and train BNNs (1 bit) and other Quantized Neural Networks (QNNs) using the familiar tf.keras interface. It introduces concepts like quantized layers and quantizers, allowing users to define how inputs and kernels are quantized. Larq is part of a broader ecosystem, including Larq Zoo for pretrained models and Larq Compute Engine for efficient deployment on mobile and edge devices.

LLM-FineTuning-Large-Language-Models is a comprehensive GitHub repository dedicated to practical techniques and projects for fine-tuning large language models (LLMs). It serves as a valuable resource for AI developers and machine learning engineers seeking to customize and enhance LLMs for specific tasks. The repository includes various fine-tuning examples for models like Llama-2, Mistral, Falcon, and CodeLLaMA, utilizing methods such as PEFT, QLoRA, GPTQ, and DPO. It also covers essential LLM concepts like 4-bit quantization, rotary embeddings, and chat templates, providing both theoretical understanding and practical implementation. This collection aims to bridge the gap between theoretical knowledge and real-world application in LLM development.

lightning-hydra-template is an open-source template designed to kickstart deep learning projects using PyTorch Lightning and Hydra. It aims to save boilerplate code, allowing users to easily add new models, datasets, tasks, and train on various accelerators like multi-GPU or TPU. The template is thoroughly commented, serving as a learning resource, and offers a collection of useful MLOps tools and code snippets for reusability. It emphasizes rapid experimentation through Hydra's command-line capabilities and minimal boilerplate via automated config instantiation. Key features include experiment tracking with tools like Tensorboard and W&B, hyperparameter search with Optuna, and continuous integration with GitHub Actions. The project structure is well-defined, supporting efficient development and debugging.

Tribuo is an open-source Java machine learning library developed by Oracle Labs' Machine Learning Research Group. It supports a wide range of prediction tasks including multi-class classification, regression, clustering, anomaly detection, and multi-label classification. The library provides its own implementations of various ML algorithms and also integrates with external tools like TensorFlow, ONNX Runtime, and XGBoost. A key feature is its use of the OLCUT configuration system, allowing repeatable model building from XML or JSON files. Tribuo emphasizes reproducibility with serializable provenance objects for models and evaluations, tracking data, transformations, and hyperparameters. It also supports exporting many models in ONNX format for deployment across different platforms.

vllm-omni is a framework designed for efficient model inference and serving of omni-modality models, building upon the foundation of vLLM. It expands support beyond text-based autoregressive generation to include text, image, video, and audio data processing. The framework also accommodates non-autoregressive architectures like Diffusion Transformers (DiT) and other parallel generation models, enabling heterogeneous outputs. Key features include state-of-the-art autoregressive support through efficient KV cache management, pipelined stage execution for high throughput, and fully disaggregated architecture with dynamic resource allocation. It offers flexibility with heterogeneous pipeline abstraction, seamless integration with Hugging Face models, and support for various parallelism techniques for distributed inference. vllm-omni also provides streaming outputs and an OpenAI-compatible API server.

Boulder AI specializes in advanced deep learning neural network cameras and edge computing devices designed for various industrial applications. Their technology enables real-time insights and critical decision-making directly at the edge, minimizing costs and maximizing accuracy by eliminating the need for extensive infrastructure. The product line includes DNN Cameras for field and traffic applications, DNN Nodes that ingest feeds from conventional IP cameras, and DNN Bullet Nano cameras for small form factor requirements. These solutions simulate human observation and decision-making, providing valuable insights for operations, compliance, and automation of routine actions across industries like smart cities, transportation, retail, and industrial sectors.

Lumana is an enterprise AI video security and intelligence platform designed to enhance physical security, safety, and operations. It leverages agentic AI to automate monitoring, deliver real-time alerts, and accelerate investigations. The platform works with existing cameras, transforming them into smart devices powered by VIA-1, Lumana’s proprietary video intelligence model. Lumana offers a comprehensive solution including Core AI engine, Cloud platform, and VMS+ smart video management software, accessible from desktop or mobile. It is camera-agnostic and provides features like face/vehicle/license plate recognition, threat detection, perimeter protection, and smart search capabilities, all backed by a lifetime hardware warranty for customers.

Vana is an open protocol designed to empower users with ownership and control over their personal data in the age of AI. It facilitates the portability and programmability of user data across various applications, moving away from traditional data extraction models. Users can connect their existing accounts, manage permissions, and control how their data is utilized within the Vana network. The platform allows developers to build applications that leverage real user context from day one, with over 1 million users already connected. Vana aims to liberate data from 'walled gardens' by providing open infrastructure for secure, user-permissioned data movement between applications, fostering a new era of data capital.

Verax AI is an enterprise-grade AI security platform designed to detect and prevent risks associated with the use of generative AI across organizations. It offers real-time visibility and enforcement of policies across employee AI usage without disrupting productivity. Key capabilities include comprehensive Shadow AI discovery to identify all generative AI tools and agents, identity-aware access control with granular, role-based policies, and content-aware DLP for real-time data leakage prevention. The platform uses AI-driven semantic analysis to inspect prompts and responses, protecting sensitive information like PII, financial records, and proprietary data. Verax AI helps CISOs, Infrastructure Leaders, and Compliance teams manage AI-driven risks, ensure policy enforcement, and maintain regulatory compliance.

Open-r1 is a fully open-source project dedicated to reproducing the DeepSeek-R1 model, providing a comprehensive framework for researchers and developers. The repository includes essential scripts for training models using techniques like Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO), as well as tools for evaluating model performance and generating synthetic data. It supports various hardware configurations and integrates with platforms like Hugging Face Hub and Weights and Biases. The project emphasizes community contribution and aims to build the missing pieces of the R1 pipeline, making advanced LLM development accessible and reproducible. It also features specialized functionalities like a code interpreter reward function for competitive programming tasks, supporting sandboxes like E2B and Morph.

Uni The Present is a premium, production-ready LMS UI starter for Next.js App Router, specifically designed for recorded courses and AI tutoring. It offers comprehensive student and instructor flows, including a student dashboard, a course viewer with video, syllabus, and notes, an AI tutor chat shell, assignments overview, messaging inbox, and a notes workspace. Instructors benefit from profiles and a studio with a grading queue. Built with Tailwind CSS 4, shadcn/ui, Framer Motion, and a Feature-Sliced architecture, it allows for easy integration with real APIs like Supabase, Postgres, and Stripe. The starter kit also includes a dedicated style guide and pre-configured metadata for quick rebranding and launch.

AI Photo is a robust application available for iOS, iPadOS, and macOS, designed to generate images from text prompts using Stable Diffusion technology. A key differentiator is its offline operation, which ensures user data privacy and prevents potential data leakage. The application supports custom CoreML models on macOS, catering to users interested in research and experimentation. Additionally, AI Photo incorporates built-in mechanisms to prevent the generation of harmful content, promoting responsible AI use. This tool is ideal for users who prioritize privacy and local processing for their image generation needs.

Diffusion Transformers (DiT) is an AI tool designed for image generation, utilizing advanced diffusion models to create visual content. It is hosted as a Hugging Face Space, making it accessible to users interested in exploring its capabilities. The project's code and model weights are licensed under CC-BY-NC, indicating its open-source nature for non-commercial use. However, the current live version on Hugging Face is encountering a runtime error, preventing immediate use. This tool is part of the broader category of AI applications focused on creative content generation.

deepseek_project is an open-source project designed to enhance locally deployed Large Language Models (LLMs) with real-time web search capabilities. Addressing the common limitation of local LLMs lacking direct internet access, this plugin enables models to fetch up-to-date information from the web, leading to more accurate and timely responses. Key features include support for multiple search engines like Google, Bing, and Baidu, the ability to summarize search results, and detailed webpage content scraping. It automatically formats search results into LLM-compatible prompts and offers a simple API for easy integration with various LLMs. The project also provides example client code, optimizes for Chinese search queries, supports real-time information retrieval, and includes a configurable web interface for parameter adjustments. Additionally, it features a WeChat Assistant for automated message replies and a Document Upload Assistant to handle file processing.

Seldon Core 2 is an MLOps and LLMOps framework designed for deploying, managing, and scaling AI systems in Kubernetes. It supports everything from singular models to complex modular and data-centric applications. With Seldon Core, users can deploy models in a standardized way across a wide range of model types, whether on-premise or in any cloud environment, ensuring production-readiness out of the box. Key features include pipelines for composable AI applications with Kafka for real-time data streaming, autoscaling based on native or custom logic, multi-model serving to optimize infrastructure costs, and experimental features like A/B tests and shadow deployments. It also allows for custom components to integrate logic, drift detection, and LLMs.

Stanford CoreNLP is a comprehensive Java-based suite of natural language analysis tools designed to process raw human language text. It can identify base forms of words, parts of speech, named entities (companies, people), normalize dates, times, and numeric quantities, and mark sentence structures. Originally developed for English, it now supports Arabic, Chinese, French, German, Hungarian, Italian, and Spanish. CoreNLP is an integrated framework, making it straightforward to apply various language analysis tools with minimal code. Its analyses serve as foundational building blocks for advanced text understanding applications, widely used in academia, industry, and government. The tools utilize rule-based, probabilistic machine learning, and deep learning components.