Research & Education

Himitsu Lab Limited is an advanced AI research organization established in 2019, dedicated to exploring intelligent systems, reasoning under uncertainty, and safe deployment of next-generation artificial intelligence. Their foundational research focuses on Native AI Intelligence, aiming for internal world understanding, uncertainty-aware reasoning, decision realism, and safe real-world action across various environments. Key research areas include World Model+ Decision Architecture, Cognitive & Reasoning Models, Swarm & Distributed Intelligence, Quantum Security & Long-Horizon Risk, and Physical & Embodied AI. The World Model+ framework is central to their approach, extending classical world models with features like uncertainty-aware reasoning, decision timing, and continuous learning to ensure safe operation under real-world constraints.

MinersAI provides an AI-driven platform specifically designed for the geoscience industry, empowering mineral exploration, mining operations, public authorities, and investors. The platform utilizes AI and machine learning to standardize, clean, and analyze complex geological data, enabling geologists to test hypotheses and uncover new opportunities. MinersAI emphasizes a 'geologists-in-the-loop' approach, ensuring that AI models are guided by expert human insight to provide meaningful and accurate results. Their solutions aim to boost success rates in mineral exploration, which traditionally has a low discovery rate, by providing high-quality, standardized data and advanced analytical tools.

Orbem leverages AI-powered MRI technology to make the invisible visible, offering fast, accurate, and non-invasive imaging solutions for various biological applications. The tool is designed to transform biological datasets into actionable intelligence, serving sectors such as poultry (in-ovo sexing, fertilization status), fruits (detecting internal defects), nuts (identifying deformation, discoloration, dehydration), and health (providing previously unattainable health insights). Orbem's mission is to build a healthier, more sustainable world by reducing waste across the food system and moving health systems towards proactive, preventive check-ups. Its non-destructive technology ensures accurate grading and reduces batch losses, contributing to global food sustainability and improved animal welfare.

OmiCure is an AI-powered platform designed to assist patients and physicians in their fight against cancer by providing personalized knowledge. The OmiCure AI analyzes a tumor's genomics and transcriptomics, comparing them against an extensive knowledge base of therapies and nutriments. This process helps identify the most suitable treatment and diet options for individual cancer patients. The platform generates a detailed report for medical teams, factoring in thousands of cancer drug-gene and nutriment-gene interactions. Developed over 10 years with leading research laboratories, OmiCure aims to keep doctors at the forefront of innovation by delivering actionable, personalized information for solid tumors at all stages.

llama.go is a pure Golang reimplementation of the popular llama.cpp framework, designed for machine learning enthusiasts and developers. It aims to provide a GGML-compatible environment for debugging and inferring large GPT models directly in Golang, offering an alternative to lower-level languages like C++. The project focuses on performance and elegance, enabling users to work with models like LLaMA-7B, 13B, 30B, and 65B. Key features include multi-threading, cross-platform compatibility (Mac, Linux, Windows), and optimizations for ARM NEON and x64 AVX2. It also supports modern GGUF V3 model format, INT8 quantization, and offers an embedded REST API for production use, allowing parallel inference with configurable pods and threads.

Liger-Kernel is an open-source collection of Triton kernels specifically engineered to optimize Large Language Model (LLM) training. Developed by LinkedIn, this tool boasts a 20% increase in multi-GPU training throughput and a 60% reduction in memory usage, enabling longer context lengths, larger batch sizes, and massive vocabularies. It offers optimized Post-Training kernels, including DPO, ORPO, CPO, and SimPO, which can deliver up to 80% memory savings for alignment and distillation tasks. Liger-Kernel is designed for ease of use, allowing users to patch Hugging Face models with a single line of code or compose custom models using its modules. It is compatible with multi-GPU setups like PyTorch FSDP, DeepSpeed, and DDP, and integrates with popular trainer frameworks such as Axolotl and Hugging Face Trainer. The kernels are exact, ensuring computational accuracy with rigorous unit tests and convergence testing.

Sravathi AI Technology Pvt Ltd pioneers AI-driven solutions for complex drug discovery and development challenges, founded in 2020. The platform leverages generative AI, predictive AI, and physics-based models to design novel compounds, identify and validate targets. It allows clients to rapidly narrow down to a few hundred prioritized compounds for chemical synthesis by considering physico-chemical, ADMET, and pharmaco-kinetic properties. This approach significantly reduces uncertainty, time, and costs in the drug discovery process, transforming chemistry with its cost-effective, rule-based, and data-driven AI platform.

DeepLearning.scala is an open-source library designed for building complex neural networks using Scala. It supports differentiable programming, allowing users to construct neural networks from mathematical formulas and calculate derivatives for weights. A key differentiator is its ability to create dynamic neural networks, where the structure can change during runtime based on Scala functions and control flows. This enables programmers to build sophisticated networks with familiar coding paradigms. The library also emphasizes functional programming, leveraging Monads and Applicative type classes for composable layers and parallel computations. DeepLearning.scala 2.0 is organized around Dependent Object Type calculus (DOT), providing mixin-able plugins for extending functionality, including algorithms, models, and hyperparameters, all with static type checking.

filter-pruning-geometric-median is an open-source implementation of the Filter Pruning via Geometric Median method for accelerating deep convolutional neural networks. Developed in PyTorch, this tool enables researchers and developers to reduce the computational cost and memory footprint of their models without significant loss in accuracy. It supports both network-level and layer-level sparsity configurations, offering flexibility in how pruning is applied. The repository provides detailed usage instructions for integration with PyTorch and NNI, along with scripts for reproducing results on datasets like ImageNet and CIFAR-10, making it a valuable resource for model compression research and application.

ASENSEI is a leading software development kit (SDK) that leverages computer vision for movement recognition and AI coaching intelligence. It enables businesses to transform workout videos into personalized experiences, offering guidance, adaptation, and rewards to boost customer acquisition, engagement, and retention in fitness. In healthcare, ASENSEI assists in onboarding patients into virtual physical therapy, providing real-time support, tracking progress, and ensuring adherence. This technology helps improve patient outcomes, reduce care costs, and facilitate scalable, high-quality treatment. ASENSEI.AI simplifies the integration of motion capture, movement recognition, and AI coaching into any hardware or software product, providing the 'brain' for AI coaches with cutting-edge computer vision and specialized large language models.

PhiFlow is an open-source simulation toolkit designed for machine learning and optimization, primarily written in Python. It offers a differentiable PDE solving framework that seamlessly integrates with popular machine learning frameworks such as NumPy, PyTorch, Jax, and TensorFlow. This integration allows users to leverage automatic differentiation for building end-to-end differentiable functions that combine learning models with physics simulations. PhiFlow supports a wide range of applications, particularly in fluid dynamics, with features like built-in PDE operations, a flexible web interface for live visualizations, and object-oriented design for extensibility. It enables reusable simulation code across different backends and dimensionalities, making it a versatile tool for researchers and developers.

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.

BharatRohan is an innovative AI-driven platform dedicated to revitalizing the agricultural value chain in India. Leveraging advanced drone-based hyperspectral remote sensing and artificial intelligence, the service provides crucial insights into land and crop health. Its core mission is to make precision agriculture and integrated pest management viable for farmers, thereby promoting sustainable farming practices. By offering data-driven recommendations, BharatRohan helps optimize resource utilization, improve crop yields, and reduce environmental impact, fostering a more resilient and efficient agricultural ecosystem.

Mozaic Earth is an AI-native platform designed for site-level nature intelligence, enabling organizations to collect and analyze habitat and biodiversity data efficiently. It facilitates collaboration with local staff and communities, allowing for data collection and analysis in a fraction of the time and cost, with full transparency. The platform features an intuitive mobile app for field data collection, making it easy for anyone to capture geolocated data, even offline. For ecologists, Mozaic Earth offers an AI-powered nature analysis platform that accelerates remote surveys and provides AI-driven recommendations, increasing capacity by 2-3X. It also generates audit-grade biodiversity risks and impact reports, combining remote sensing insights with ground-truth data to quantify and track environmental impact.

NatureDots offers AquaNurch Digital Twin, a foundational AI layer for water and aquatic environments, integrating data layers with user-specific functional layers. This system constructs operational Digital Twins that interpret complex ecosystem dynamics at a geo-precise scale, delivering timely insights and predictive foresight. Deployed across 275,000 hectares, the platform significantly reduces operational effort by 20x and lowers both CAPEX and OPEX. The suite includes Twinity for water industry users needing predictive intelligence beyond monitoring, Twingills for next-gen water and fish intelligence in aquaculture, and Twinsfera for monitoring ecosystem changes, biodiversity risks, and climate shifts. NatureDots aims to build 10 million hectares of climate-resilient waterscapes.

The RoboCup Federation is an international scientific initiative dedicated to advancing the state of the art in intelligent robots and artificial intelligence. Established in 1997, its original mission was to develop a team of robots capable of defeating human soccer champions by 2050. The federation organizes various competitive leagues, including RoboCupSoccer, RoboCupRescue, RoboCup@Home, RoboCupIndustrial, and RoboCupJunior, fostering research and development in diverse robotic applications. Beyond competitions, RoboCup hosts symposiums, publishes research papers, and offers awards to recognize significant contributions to the field. It serves as a platform for international collaboration, bringing together researchers, students, and enthusiasts to push the boundaries of robotics and AI.

torch.rb provides deep learning capabilities for Ruby developers, leveraging the power of LibTorch. It allows users to create and manipulate tensors, perform various operations, and build neural networks directly within the Ruby environment. The library closely follows the PyTorch API, with minor adjustments to be more Ruby-like, making it easier for developers familiar with PyTorch to transition. It supports tasks such as image classification, collaborative filtering, and generative adversarial networks, and integrates with TorchVision, TorchText, and TorchAudio for specialized computer vision, NLP, and audio tasks. Performance can be significantly enhanced on GPUs, with support for CUDA on Linux and Metal Performance Shaders (MPS) on Mac.

Vivent Biosignals offers an advanced solution for agriculture by leveraging biosensors and AI algorithms to decode real-time plant biosignals. This technology allows farmers, indoor growers, and academics to make plant-driven decisions, optimizing growing conditions, detecting stress, diseases, and pests before visual symptoms appear, and accelerating the development of new crop inputs. The system provides insights into how crops respond to their environment, leading to improved yields, reduced inputs, and more sustainable practices. Vivent Biosignals is a B Corp-certified company, committed to social and environmental standards, and supports various agricultural settings including fields, greenhouses, and indoor farms.

Augment Technologies offers the AI-driven Muckpile Block Model™, a world-leading 3D blast movement solution for open pit mining. Powered by artificial intelligence and a physics engine, it accurately predicts ore material landing spots minutes after a blast. This technology helps increase metal ore recovery by up to 5%, minimizes ore loss and dilution, and can generate tens of millions of dollars in additional revenue per mine annually. The software-only solution integrates seamlessly with existing workflows and general mining software like Datamine™, improving site safety by reducing the need for personnel on drill pads and muckpiles. It provides precise information for geology, drill and blast, and management teams to optimize operations and improve NPV.

Variational AI leverages advanced generative AI through its Enki™ platform to revolutionize early-stage drug discovery. Enki™ generates novel, synthesis-ready, lead-like compounds tailored to specific target product profiles, effectively eliminating the need for traditional hit identification and hit-to-lead phases. This allows biopharmaceutical partners to move directly into lead optimization with structures not discoverable by conventional methods. The platform optimizes across more than 50 parameters, including potency, selectivity, ADMET, and synthetic feasibility, for 760 pre-trained drug targets. By designing de novo molecular structures, Variational AI aims to provide better starting points for drug programs, leading to fewer costly iterations, faster timelines, and a higher probability of success.

NVIDIA Model Optimizer is an open-source library designed to accelerate deep learning models through various state-of-the-art optimization techniques. It supports quantization, pruning, distillation, speculative decoding, and sparsity to compress models and enhance inference speed. The tool accepts Hugging Face, PyTorch, or ONNX models as input and provides Python APIs for composing optimization techniques. Optimized checkpoints can be seamlessly exported for deployment in frameworks like SGLang, TensorRT-LLM, TensorRT, and vLLM, making it a crucial component within the NVIDIA AI software ecosystem for efficient model deployment.

Purdue AI Racing is a program at Purdue University dedicated to fostering research and education in artificial intelligence. The initiative provides a platform for students and faculty to delve into various AI applications, particularly within the fields of engineering and robotics. It offers essential resources for the research and development of autonomous vehicle technology, contributing significantly to the university's broader mission of innovation in science and technology. The program aims to push the boundaries of AI knowledge and practical implementation through academic exploration and hands-on projects.

YOLO-World is a cutting-edge, real-time open-vocabulary object detector, developed by AILab-CVC and accepted by CVPR 2024. This open-source project provides PyTorch implementation, pre-trained weights, and code for both pre-training and fine-tuning. It excels in open-vocabulary detection and grounding, allowing users to detect objects without pre-defined categories. The tool is pre-trained on large-scale datasets, including detection, grounding, and image-text datasets. YOLO-World introduces a prompt-then-detect paradigm for efficient user-vocabulary inference, re-parameterizing vocabulary embeddings into the model for superior inference speed. It supports various fine-tuning recipes, including normal fine-tuning, prompt tuning, and reparameterized fine-tuning, making it adaptable for custom datasets and specific domains. Additionally, it offers deployment options for ONNX, TFLite, and INT8 Quantization.

Brainamics provides the first and only objective playtesting solution for the gaming industry. By leveraging cutting-edge neurotechnology and research-grade machine learning, the platform helps game developers understand user psychology and optimize gameplay. This innovative approach allows developers to gain deep insights into player engagement, emotional responses, and cognitive load, which are crucial for refining game mechanics and overall user experience. Brainamics aims to elevate games to the next level, enabling studios to get a significant advantage in the increasingly competitive gaming market by making data-driven decisions based on real-time neural feedback.