At Quark AI, we believe AI should move forward, fast.
Not slowed by overregulation or endless safety bureaucracy. Just the right guardrails. rapid progress towards AGI.
Explore ProjectsQuark AI is a research and development company dedicated to pushing the boundaries of artificial intelligence. Founded by Dheeraj Kumar, we focused on making ai reach its full potential.
EasiScriptX (ESX) is a high-performance domain-specific language (DSL) designed for AI/ML workflows. It simplifies defining models, datasets, and training routines with a declarative syntax. It's human-readable , efficient, reduces development time..
Key features include advanced tensor operations (matrix multiplication, convolution, pooling, normalization), distributed training with multi-GPU (NCCL) and multi-node (MPI) support, loading pretrained models via ONNX Runtime or PyTorch, custom loss functions and optimizers, profiling and debugging tools, autonomic optimization (model compression, multi-agent tuning), and extensibility for new layers and operations.
It supports parameter-efficient fine-tuning (LoRA), high-performance attention (FlashAttention-2), mixed-precision training (BF16/FP16), pipeline parallelism, domain adaptation, energy-aware and heterogeneous scheduling, interoperability with PyTorch, TensorFlow, and JAX via ONNX, efficient data pipelines, experiment tracking with MLflow, and optimization for both high-end and low-end hardware.
Built with C++20, Boost, Eigen, ONNX Runtime, OpenMPI, and optional CUDA. Currently at version 1.0, licensed under MIT. Ideal for hobbyists and enterprise workflows, reducing memory and energy usage.
View on GitHubCongming AI is an experimental, modular, neuro-symbolic Artificial Intelligence Framework with a Multi Modular Multi Agent System, focused on advanced reasoning capabilities.
It had successfuly modified some files of code by itself (about eight files) , and its capable of choosing an random topic and learning about it via internet (which we call ingest, it stores data in its database) and uses information whenever needed.
Benchmarking shows Congming-2 with 43% pixel accuracy and ~30% task success on ARC-1, and Congming-3 with ~57% average pixel accuracy and 18% task success on ARC-AGI2 across ~200 tasks.
Technologies include external API calls and local Hugging Face models.
One of our main products, aiming for rapid, autonomous ai development with minimal restrictions.
View on GitHubThinking of ai beyond traditional transformer architecture, we're desgining Synapse ai model entierly from scratch, for reasoning, not just token prediction.
In development
One of our long term goal, Custom AI chips designed for our Synapse-AI architecture, optimized for efficiency, power consumption, and high-performance computing in AI tasks.
Currenlty very early stage of design.