Artificial Intelligence

We are not obsessed with Scale we are obsessed with Intelligence by maximizing Logic, Math, Causality; by understanding and explaining, using Self Learning, and minimizing Energy, Compute and Training Data

OUR FOCUS

Multi-Task, Multi-Modal, Generalizable, Un/Semi Supervised, Reliable, 3D, Efficiency

01.

Intelligence that Understands and Explains, is Reliable and Generalizable

We understand something when we self learn reusable and repurposeable internal representations.
The same is used by our AI systems for inference and explaining the results. This philosophy underpins our entire approach to AI

03.

Natural Language Understanding & Processing

Our NLU models are not Stochastic Parrots, actually they are very far from it. While we use standard models of AI like transformers for Natural Language Tasks across 40+ human languages. With Text and Speech Modalities. For Business Applications like Q&A, Search, Query, Conversations, Translations etc. We strive to create and leverage knowledge from the inputs. To make intelligent conclusions which are paramount for business success.

04.

Vision

We combine 3D Graphics, Image Processing, Graph Networks to create myriad applications like 2D to 3D Projections into the Metaverse, Collaborative VR Tools, Recognition and Classification in 2/3D, Multi-Modal Search. Generative and Parametric Models in AI. We dabble regularly with Probabilistic Computing, Logic Networks and Spiking Neural Networks for our customer projects.

02.

New product design & development

Being able to do Maths, Logic and Causality is the final frontier in AI. We are in the business of building ‘The Intelligence’ in systems. Just because you can maximize something we shouldn’t, it is not that it gives depreciating returns but in AI its retrograde to maximize training data, model size/scale etc. The best AI is one which minimizes Energy Consumption, Training Data, Compute, #Parameters, Scale. We don’t salivate about training trillion parameter neural networks.

Case Study

A product for
predicting marketing ROI at a Fortune 500 Co.

We have built an internal tool that reduces the manual efforts put by
data scientists in a fortune 500 Co. by a whopping 60%.

How it works

We constantly keep abreast of the top innovations in A.I. from Industry and Academia across The World. We focus on Open Source and try to stay away from Vendor Lockins. For 50% of the 'Simple' use cases we apply off-the-shelf models or fine tune them with context specific training data. For other use cases we start ab initio by creating datasets, labeling them, spiking models, hyper-parameter tuning, tests and deployment.

A.I. Operations

We have extended our DevOps CI/CD pipelines to include DVC/MLOps, SecOps and FinOps, and Continuous Learning for AI/ML Models. To apply engineering rigor to A.I. development.

Design Thinking in A.I.

Prototype-Test-Iterate Loops using Levers like – UX, Bias, Adversarial Attacks, Training Data Recovery Attacks, Explainability, Reliability, Accuracy, Efficiency, Costs of Failures, Privacy, Ethics, Law, Energy, Compute, Memory, Latency, Network, Costs, Carbon Footprint, and Various Tradeoffs.

Federated A.I.

Our early efforts for XXL models were around distributed training and inference on shared lowlatency infrastructures. But with new regulations on Privacy we have adopted Privacy Preserving Federated Learning Architectures. We have also spiked Public Compute Grid Architectures using WebAssemblies.

Our USP

While we have trained models with billions of parameters and aspire to touch a trillion parameter models very soon, given the right customer requirement. But we are very conscious of Training and Inference Memory, Compute, Energy, Efficiency, Carbon Requirements and Cost Minimization. We have deployed production grade models on the cloud and the edge alike with cpu, gpu, npu, mW/100kb devices. With Serverless, Browser, OCI, WebAssembly – packaging, deployment, auto scaling/monitoring. We quantize using [X/Q]NNPACK and acceptable tradeoffs.

Our Accolades 