AI Agents & Automation

Falcon-H1-Tiny offers a series of compact language models designed to push the boundaries of AI capabilities at a small scale. These models are available on Hugging Face Spaces and are ideal for research and experimentation. Users can input prompts and receive generated responses from these lightweight but capable AI models, making them suitable for various applications including research paper analysis, data visualization, and the development of small-scale AI applications. The focus on models with 100 million parameters or less makes them particularly efficient and accessible for developers and researchers working with limited resources.

SandboxAQ is a B2B company that utilizes the compound effects of AI and advanced computing to tackle significant societal challenges. Their Large Quantitative Models (LQMs) are applied across diverse fields, including AI drug discovery, new chemicals and materials innovation, cybersecurity, navigation, and medical diagnostics. The platform offers technologies such as AI simulation, cryptography management for enhanced cybersecurity, and AI sensing for global organizations. SandboxAQ's approach focuses on quantitative AI, grounded in physics and chemistry, to provide real-world solutions with clear outputs and reduced uncertainty, making it scalable and proven for high-stakes domains.

Tiktoken-go is a Go port of OpenAI's tiktoken library, designed for efficient Byte Pair Encoding (BPE) tokenization. This tool allows Go developers to seamlessly integrate tokenization capabilities into their applications, particularly when working with OpenAI's various language models like GPT-3.5, GPT-4, and embedding models. It features a cache mechanism, similar to the original Python library, which can be configured via the TIKTOKEN_CACHE_DIR environment variable to store token dictionaries and avoid repeated downloads. For scenarios requiring offline operation or custom dictionary loading, Tiktoken-go supports alternative BPE loaders, including an offline loader that uses embedded files. The library also provides utility functions for counting tokens in chat API calls, adapting to different model versions and their specific token calculation rules.

3Plus1.ai specializes in delivering value creation through expert AI implementation tailored for private equity firms. The platform focuses on enhancing due diligence processes, optimizing portfolio company operations, expanding total addressable markets (TAM), and ultimately maximizing exit multiples. With a proven track record of delivering over $100M in value creation across more than 10 portfolio companies, 3Plus1.ai is trusted by top-tier PE funds. The service is available to select partners, requiring existing partners to enter their email for access or new partners to request access. It positions itself as a trusted partner in AI solutions, creating smarter systems for smarter businesses.

Rabit is a lightweight, open-source library designed to provide a fault-tolerant Allreduce and Broadcast interface, primarily for distributed machine learning applications. It enables easy implementation of distributed machine learning programs that benefit from the Allreduce abstraction. Key features include portability, allowing it to run on various platforms like Yarn (Hadoop) and MPI with the same codebase, and scalability due to its efficient communication model. Rabit also offers reliability through synchronous function calls for model and result recovery, and supports operations before checkpoint loading. While recent developments have moved to dmlc/xgboost, Rabit remains a foundational component for distributed XGBoost.

Trigo is a Vision AI company specializing in retail solutions, offering advanced technology for loss prevention, frictionless autonomous stores, and comprehensive retail intelligence. Their proprietary platform processes over 60 million shopping activities annually with high accuracy, utilizing 100% non-biometric data to ensure privacy and compliance. This allows retailers to gain real-time alerts on theft, analyze interactions, and classify products, all while maintaining customer trust. Trigo's solutions help retailers modernize operations, gain data-driven insights, and enhance the shopping experience, addressing key challenges in the evolving retail landscape.

agenta is an open-source LLMOps platform designed to accelerate the development of reliable LLM applications. It offers a comprehensive suite of tools for prompt management, evaluation, and observability, all in one place. Key features include an interactive LLM playground for side-by-side prompt comparison, multi-model support, and version control for prompts and configurations. For evaluation, Agenta provides flexible test set creation, pre-built and custom evaluators, and human feedback integration. The platform also offers robust observability with cost and performance tracking, detailed LLM tracing, and OpenTelemetry native compatibility. It's ideal for teams looking to streamline their LLM development workflow from experimentation to production.

FOURIER-Robotics GR-2 is a cutting-edge humanoid robot designed to push the boundaries of agility, precision, and perception. Built upon customer feedback, GR-2 integrates advanced hardware, design, and software enhancements. Its next-level hardware design includes integrated cabling for power and communication, resulting in concealed wires and a more compact form factor. The improved joint configuration simplifies debugging, reduces manufacturing costs, and enhances the robot's ability to transition from AI simulation to real-world applications. GR-2 features 12-DoF dexterous hands, doubling the dexterity of previous models, and is equipped with six array-type tactile sensors for real-time force sensing and object manipulation. Powered by seven types of distinct FSA actuators, including FSA 2.0 with peak torques exceeding 380 N.m, GR-2 achieves dynamic mobility and precise movements. The Fourier Toolkit provides developers with an upgraded Software Development Kit, offering easy access to pre-optimized modules via intuitive APIs and supporting frameworks like NVIDIA Isaac Lab, ROS, and Mujoco.

Frontier Foundry specializes in developing and deploying secure, enterprise-grade AI systems tailored for highly regulated environments such as banking, asset management, healthcare, and government. Their solutions are built to operate within strict regulatory, operational, and security constraints, ensuring data never leaves the client's network. Key offerings include Kundi, an ensemble reasoning engine for institutional-scale data analysis, and Limni, a command and control platform for orchestrating AI models within secure infrastructures. Frontier Foundry focuses on delivering measurable outcomes, providing end-to-end audit trails, and ensuring deterministic, reproducible AI outputs suitable for critical decision-making and regulatory examination.

lite.ai.toolkit is a comprehensive C++ AI toolkit offering over 100 pre-trained AI models for a wide range of computer vision tasks. It supports multiple inference engines including MNN, ONNX Runtime (ORT), and TensorRT (TRT), allowing for flexible deployment across different platforms like Linux, macOS, Windows, and Android. The toolkit includes models for object detection (YOLO series, SSD, EfficientDet), face detection and recognition (RetinaFace, ArcFace), segmentation (MODNet, RobustVideoMatting), and other advanced functionalities like Stable Diffusion and Face Fusion. It emphasizes simplicity and user-friendliness with consistent syntax, minimal dependencies, and detailed build instructions for easy integration into C++ projects.

LatentMAS is a multi-agent reasoning framework designed to enhance the efficiency and stability of multi-agent systems. Unlike traditional methods that rely on lengthy textual reasoning traces, LatentMAS facilitates agent collaboration by passing latent thoughts directly through their working memory within the model's latent space. This innovative approach significantly reduces token usage by 50-80% and achieves major wall-clock speedups of 3-7 times compared to standard Text-MAS or chain-of-thought baselines. The framework is compatible with any HuggingFace model and optionally supports vLLM backends for faster inference. It also features training-free latent-space alignment for stable generation, making it a general and powerful technique for developing advanced multi-agent AI applications.

Sahaj Software is an artisanal technology services company focused on delivering purpose-built solutions through intelligent engineering. They specialize in AI, ML, data engineering, and platform engineering, helping organizations achieve data-led transformation. Their approach emphasizes simplicity, first principles thinking, and lean cohesive teams to solve complex problems. Sahaj offers technology advisory services, including tech due diligence and assessment, to provide informed decision-making and better risk management. They are committed to full knowledge transfer, ensuring clients are not dependent on Sahaj post-implementation. The company's ethos is rooted in trust, respect, curiosity, and craftsmanship, aiming to inspire brilliance and reduce exploitation.

Marian is an efficient open-source Neural Machine Translation framework implemented in pure C++ with minimal dependencies. It is designed for high performance, supporting fast multi-GPU training and GPU/CPU translation. The framework incorporates state-of-the-art NMT architectures, including deep RNN and transformer models, making it suitable for advanced machine translation research and development. Marian is released under a permissive MIT open-source license, encouraging broad adoption and contribution. Its focus on efficiency and C++ implementation provides a robust foundation for building and deploying neural machine translation systems.

node-tensorflow is an open-source project offering a NodeJS API for Google's powerful machine learning library, TensorFlow. This initiative focuses on making TensorFlow's capabilities readily accessible to JavaScript developers within the NodeJS environment, prioritizing both performance and stability. Currently in its early design stages, the project actively seeks contributions, particularly from individuals with C++ knowledge, to accelerate its development. It leverages SWIG for interfacing the C++ core API with Node.js bindings, with a roadmap that includes integrating Python API features like Optimizers and Tensor Transformations. The goal is to evolve into a robust Node.js API for end-users, enabling them to build and control TensorFlow computation graphs directly from Node.js.

ModelingToolkit.jl is a high-performance symbolic-numeric computation framework designed for scientific computing and scientific machine learning within the Julia ecosystem. It allows users to define models at a high level, enabling symbolic preprocessing for analysis and enhancement. The tool can automatically generate optimized functions for model components, such as Jacobians and Hessians, and automatically sparsify and parallelize computations. It also applies automatic transformations, like index reduction, to simplify models for numerical solvers. ModelingToolkit.jl supports composing multiple ODE subsystems and simulating complex Differential-Algebraic Equations (DAEs), making it a powerful tool for advanced scientific modeling and simulation.

DeepLearningKit provides an open-source deep learning framework specifically designed for Apple's platforms, including iOS, OS X, and tvOS. Developed using Metal and Swift, it offers optimized performance for deep learning applications running on Apple devices. The framework supports various deep learning functionalities, enabling developers to integrate AI capabilities into their mobile, desktop, and TV applications. It includes resources like video tutorials for getting started on different Apple operating systems and a publication detailing its architecture and development. Being open source under the Apache 2.0 Licence, it encourages community contributions and provides a foundation for building custom deep learning solutions within the Apple ecosystem.

Digital Software Labs specializes in providing purpose-driven digital solutions, including mobile app development, web applications, AI solutions, and cloud services. Their expertise spans Flutter, Android, iOS, and React Native app development, alongside comprehensive web development covering frontend, backend, and e-commerce. They also offer digital marketing services like SEO and social media marketing. The company focuses on creating custom software, enterprise solutions, and AI-driven applications, ensuring technical excellence from brainstorming to launch. They cater to startups, growing businesses, and enterprises, offering services like MVP development, platform optimization, cloud services, and ongoing support. Their approach emphasizes meticulous planning, user-centric design, security, and long-term scalability.

DI-engine is a generalized decision intelligence engine built for PyTorch and JAX, offering a comprehensive framework for reinforcement learning. It features python-first and asynchronous-native task and middleware abstractions, integrating key decision-making concepts like Env, Policy, and Model. The framework supports a wide array of deep reinforcement learning algorithms, including DQN, PPO, SAC, and multi-agent, imitation, offline, and model-based RL. Beyond algorithms, DI-engine aims to standardize decision intelligence environments and applications, catering to academic research and prototype development. It also includes highly re-usable modules for RL optimization, PyTorch utilities, and system optimizations for efficient large-scale RL training.

Float16 is a comprehensive GPU management platform designed for deploying, managing, and scaling AI models. It offers a full spectrum of services including AI-as-a-Service (AaaS) for instant access to ready-to-use AI models without coding, Platform-as-a-Service (PaaS) for flexible resource allocation, and Infrastructure-as-a-Service (IaaS) for bare-metal GPU instances. The platform emphasizes ease of use with one-click deployment, significantly reducing setup time from weeks to minutes. Float16 provides dedicated and isolated GPU resources, ensuring zero interference and optimal performance for workloads. It features a credit-based quota system for flexible GPU utilization, eliminating waste from fixed time slots. Supported by NVIDIA Inception Program, Float16 is ideal for ML engineers, data scientists, software developers, and researchers seeking efficient and scalable GPU solutions.

Neuropod is a library designed to offer a uniform interface for running deep learning models across various frameworks, including TensorFlow, PyTorch, TorchScript, Keras, and Ludwig. It aims to simplify the productionization of deep learning models, enabling researchers and developers to build models in their framework of choice without being constrained by deployment complexities. A key benefit is framework-agnostic inference code, allowing easy switching between deep learning frameworks without altering runtime code. Neuropod also supports defining a problem API, which helps in building generic tools, pipelines, and comparing models solving the same problem, even if they originate from different frameworks. It supports both C++ and Python, offers efficient zero-copy operations, and ensures model isolation with out-of-process execution.

EasyML is a general-purpose dataflow-based system designed to ease the process of applying machine learning algorithms to real-world tasks, especially on distributed platforms such as Hadoop and Spark. It formulates learning tasks as directed acyclic graphs (DAGs), where each node represents an operation or algorithm. The system includes a distributed machine learning library with algorithms for pre/post-processing, data transformation, feature generation, and performance evaluation, primarily based on Spark. A GUI-based studio allows users to create, configure, submit, monitor, and share machine learning processes using a drag-and-drop interface. EasyML also offers a cloud service for executing tasks, scheduling nodes automatically on Linux, Spark, or Map-Reduce based on their implementation. Users can upload their own algorithm packages and datasets.

drl_grasping is an open-source project focused on advancing robotic manipulation through deep reinforcement learning. It enables robots to acquire robust grasping policies for diverse objects using compact 3D observations in the form of octrees. The project emphasizes sim-to-real transfer, allowing policies trained in simulation to be evaluated on real robots with zero-shot transfer. It includes multiple RL environments for robotic manipulation, supporting continuous actions in Cartesian space and various observation variants like RGB images, depth maps, and octrees. The framework is compatible with Gym API and has been tested with end-to-end model-free actor-critic algorithms like TD3, SAC, and TQC, with a setup for model-based algorithms also provided.

NeuroBlock is an AI laboratory dedicated to enhancing AI models through the use of high-quality datasets. The platform provides comprehensive enterprise AI consulting services, assisting businesses in integrating and optimizing AI solutions. A key offering includes local and private AI integrations, ensuring data privacy and tailored performance for specific organizational needs. Additionally, NeuroBlock features an OpenData platform, designed to facilitate AI model training by providing access to diverse and curated datasets. The company also develops lead generation tools, leveraging AI to identify and engage potential customers. NeuroBlock aims to deliver AI solutions that are efficient, secure, and customized to client requirements.

facexlib is an open-source library designed to provide ready-to-use face-related functions, leveraging current state-of-the-art open-source methods. It primarily offers PyTorch reference codes for various face processing tasks, including detection, alignment, recognition, parsing, matting, headpose estimation, and tracking. While it provides a collection of these algorithms, users are directed to the original repositories for training or fine-tuning. The library simplifies the integration of advanced face processing techniques into existing projects, making it a valuable resource for developers and researchers working with facial data. It is released under the MIT license, with individual components referencing their original licenses.