EKS, LLC offers AI/ML services based on its patent-pending and proprietary technologies for NLP, image, voice and video based applications. We can develop highly performant, custom, transformer-based multi-modal models for you. If you already have a model we can reduce its parameter counts, in most cases, so that your model trains faster and cheaper, produces results that are better and reduces total ownership costs therby making it greener.

One of our patent-pending methods recast the 200-year-old linear regression model to use bounded linear computations. In anothor method we extend this approach to apply to anywhere a linear operator is used. In addition, we have a method that allows for unrestrained rotations in n-dimensional space.

Our method doesn't require batch normalization or layer normalization. Both batch norm and layer norm project the product of weight vectors with input vectors onto a unit ball based on the mean and variance of the products and then use scaling operators to come off the unit ball. Projecting onto a unit ball looses information that is present in the relative magnitudes of the input vectors.

The upshot is that our method passes information better between layers, has strong convergence and stability properties.

See our website for further details.

See our recent article on Gas-Efficient Methods for Constant-Function Automated Market Makers in DeFi on LinkedIn.

We have made available a starter kit for universal web3 app development using traditional and blockchain based backends called web3-sveltekit-bundle

We will be making available Uniswap V4 hooks, our protocols and our governance token shortly.