Frontier models generate answers. MarutAI is the frontier model for understanding how the real world fails.
We turn near-misses, drift, operator error, and system limits into customer-specific failure knowledge that determines what AI is allowed to do.
That is how teams can move faster on hard problems without asking AI to guess. The system learns where the operation breaks, acts inside defined boundaries, and improves with every run.
The Business
An organization's advantage begins with its own data and operating history. We help turn the outcomes, exceptions, near-misses, and decisions that shaped the business into a reliable foundation for AI.
From that foundation, we build models around the decisions and operations that matter to the organization. Our platform, Manifest, creates the infrastructure and harness to put those models to work. Teams use it to build workflows and agents, connect tools and systems, define guardrails, and manage security, compliance, and oversight.
Together, this turns a customer's failure knowledge into a proprietary advantage that compounds with use. The system can take on more of the work without forcing the customer to choose between autonomy and control.
The Science
Together, Catalyst and Verified Adaptive Twin Systems (VATS) create a continuous loop between what the system knows and what it is allowed to do.
Catalyst identifies gaps in the available data and determines what should be learned next. VATS uses that growing knowledge to construct decision models, then tests them against an operational world model and synthetic edge cases that historical data may never contain.
Each model carries a defined range of action, the assumptions behind it, and the evidence used to build its agent harness. New data feeds the loop again, allowing capability and control to improve together.
Institutions
University of Oxford
Columbia University
Yale University
UC San Diego
University of Waterloo
UNSW Sydney
IIT HyderabadUniversity of OxfordColumbia UniversityYale UniversityUC San DiegoUniversity of WaterlooUNSW SydneyIIT HyderabadThe Platform
Manifest is the platform that puts those models into operation. It carries each model's boundaries into the workflows and agents that use it, connects them to the customer's data and tools, and controls how actions are approved, blocked, or escalated.
Customers can run MarutAI models built with Catalyst and VATS alongside models from more than 90 commercial providers. Manifest manages routing, credentials, access controls, fallbacks, and monitoring across the deployment.
Every governed action can be traced from the data and model that informed it through the approval, constraint decision, and outcome. Teams gain the evidence to expand autonomy without losing control.
Build, deploy, and govern across leading foundation models
Anthropic
OpenAI
Google
Meta
Microsoft
NVIDIA
DeepSeek
QwenAnthropicOpenAIGoogleMetaMicrosoftNVIDIADeepSeekQwenThe Outcomes
A MarutAI deployment turns the customer's data and operating knowledge into a capability competitors cannot buy off the shelf. It improves as the operation generates new evidence, so the value compounds rather than resetting with each project.
The practical payoff is better decisions with less manual intervention and a clear path to autonomous operation. Customers can expand what the system does while keeping control of where it stops.
The Team
MarutAI was founded by Preetam Dutta and Rob Bathurst, after careers building and securing AI and data systems for DARPA, the Department of Defense, Mayo Clinic, Bloomberg, Cylance, and BlackBerry.
The team combines original research with hands-on deployment. We hire people who can turn a formal idea into working infrastructure and stay accountable for how it behaves in the real world.
MarutAI is a member of the NVIDIA Inception Program and Google for Startups. We work with academic partners and contribute to open source.
