Research.

Infrastructure.

AI requires robust infrastructure. Our research in this layer is about creating robust frameworks and optimizations that will allow us to squeeze more performance and output from available infrastructure. Our research currently focuses on optimizing training and inference. We are also working on a high performance storage and data retrieval system to manage data across applications. We have developed capabilities for on-prem as well as cloud deployments of the AI systems we develop.

Data.

Data is the most crucial part of an AI system. And we have developed proprietary data pipelines that allow us to have access to specific and high quality data. Our in-house Data Fabric pipeline allows us to generate task and system specific data for the end applications. We manage the whole data acquisition, cleaning and labelling workflow in-house due to the sensitive nature of some of the data we handle.

Models.

Our work mainly focuses on adaptation training of models. We use primarily post-training techniques to develop models suited for specific domains and tasks. Our approach to adaptation training of models is chosen by thinking backwards from what the AI system will need to do or tasks it will need to perform, And then designing the model training for that.
We also are currently working on reducing the costs needed to pre-train small sized models from scratch.

Inference.

For AI to be implemented across the board and widely like electricity or internet is, it needs to be as available and affordable as those utilities as well. And the most crucial lever we can pull for that outcome is to optimize and minimize the cost of serving these models. Our main at the inference layer is to keep reducing cost of inference to the point a freemium model will make sense. Our efforts so far have resulted in 1 million token costs below 20 cents. We aim to drive costs down even more and build AI systems that are cheap and scalable for all the applications to be built.

Applications.

We think the current landscape is focusing too much on the models and systems and not enough on the downstream capabilities AI enables. All the intelligence available is pointless if its not useful for solving problems and improving work. We focus heavily on applications, agents and workflows that can create real world value. We are currently focusing on use cases in manufacturing, finance and general business use cases.